Statistical Forecasting

Statistical Forecasting
Statistical Forecasting
Warren Gilchrist
n
Head of Department of Mathematics and Statistics,
Sheffield City Polytechnic
A Wiley—Interscience Publication
JOHN WILEY & SONS

Chichester • New York • Brisbane • Toronto
Copyright ©1976 by John Wiley & Sons, Ltd.
Reprinted September 1977

Reprinted April 1978
All rights reserved.
No part of this book may be reproduced by any means, nor

translated, nor transmitted into a machine language without
the written permission of the publisher.
Library of Congress Cataloging in Publication Data:
Gilchrist, Warren Garland, 1932—

Statistical forecasting.
‘A Wiley—Interscience publication.’
1. Time-series analysis. 2. Prediction theory.
I. Title.
QA280.G54 519.5'4 76-13504
ISBN 0 471 99402 2 (Cloth)

ISBN 0 471 99403 0 (Paper)
Typeset in IBM Century by

Preface Ltd, Salisbury, Wilts
and printed in Great Britain by
The Pitman Press, Bath
Preface
From the dawn of recorded history, and probably before, man has
sought to forecast the future. Indeed, the ability to foresee the
consequences of actions and events is one of the defining properties of
‘mind’. Where the phenomenon being forecast was a purely physical
one, such as the occurrence of midsummer’s day or of an eclipse, man
was able from very early times to obtain very accurate forecasts. Initially
such forecasts were derived on a purely empirical basis. Methods were
found that worked without any basic understanding as to why they
worked. Later, as greater understanding of the phenomena developed
theoretical models were developed which enabled forecasts to be
obtained on a more reasoned basis. During the last century interest
focused on a number of phenomena, such as economic variation and
sunspot cycles, where
(a) there were available series of observations taken over a period of
time, called time-series, upon which forecasts could be based;
(b) purely mathematical rules were found to be inadequate to describe
the phenomena since they involved features of a chance nature.
Over the last fifty years approaches to this type of problem have been
developed which seek to allow for the influence of chance and for the
unknown complexity of these situations. Thus, forecasting methods
were developed which were essentially statistical in nature. These were
based on using statistical techniques to fit models of a wide variety of
types to past time-series data and on forecasting the future from these
models.
The aim of this book is to provide the reader with an understanding
of the methods and practice of statistical forecasting. The structure of
the book is based on the structure of the activities involved in the
practice of statistical forecasting. Figure P.l summarizes this structure.
Part I of the book provides a general introduction to some of the
basic concepts and terms in statistical forecasting. Part II deals with the
development of statistical forecasting methods applicable to each of a
variety of different situations. The approach to classifying this part into
chapters has been to classify the situations; e.g. one chapter considers
v
VI \
Part I
Figure P.l The structure of the book and of statistical

forecasting
trending time series, another seasonally varying time series. We then

consider the various methods that might be used in each situation. In
practice there is considerably more to statistical forecasting than putting
data into a formula to obtain a forecast. Part III deals with these
further facets. In Chapter 13 we look at the data that is needed for
forecasting, where it might come from and how it may be treated.
Chapter 14 looks at ways in which the forecasting methods of Part II
may be adapted to apply to a wider range of applications and to give a
better quality of forecast. In Chapter 15 we examine a wide range of
methods for analysing and comparing forecasting methods. Chapter 16
has the same basic aim, but is conerned more with the routine control
of an operational forecasting system. In Chapter 17 we use some of the
properties of forecasts, that might be identified using the methods of
Chapters 15 and 16, to create even better forecasts. Always we seek to
bear in mind that forecasts are produced for practical purposes that
may influence the way we obtain and use our forecast in the first place.
Vll
This interrelation between forecast and application is explored in

Chapter 18.
The general aim of the book is to give a systematic account of the
methods of statistical forecasting in common use and of the practical
aspects of their use; consequently, the level of mathematics and
statistics assumed of the reader has been kept to that which the writer’s
teaching experience suggests as appropriate to people whose interest is
in obtaining forecasts for practical purposes. This is mainly limited to
the basic ideas of statistics and the ability to follow simple algebraic
arguments. Where more than this is used, the section is marked with an
* and can be omitted without detriment to the understanding of later
sections. Where certain ideas are used that might be unfamiliar to a
significant proportion of readers, these are discussed in the appendices.
Wherever possible, the discussion of each forecasting method is
developed and illustrated by example in such a way that the reader may
then be able to use the method himself. There are, however, a number
of methods whose mathematical development has been regarded as
beyond the scope of this book, but computer programs are commonly
available for their use. The author’s aim in these cases has therefore
been to discuss and illustrate the basic concepts of the methods. Thus
the reader should be able to use such computer programs with
understanding of their principles though not of their technical details.
A short list of relevant references which are referred to in the text
are given at the end of each chapter.
Acknowledgements
I would like to express my thanks to the many people who have

discussed forecasting problems with me over the years and those who
have given me comments on early drafts of this book. In particular, I
would like to thank N. Booth, O. Oliver, V. G. Gilchrist and also
numerous students who have helped me develop suitable approaches
to many of the topics in this book. Finally, I would like to thank the
Social Science Research Council for their support of work on the
analysis of forecasting methods that underlies, particularly, Chapter 15.
W. G. GILCHRIST
N.
.
Contents
Notation xiii
PART I PRELIMINARIES
Chapter 1. An introduction to forecasting 3

1.1 Everybody forecasts 3
1.2 Methods of forecasting 4
1.3 The basis of forecasting 5
Chapter 2. Models 12
2.1 Scientific forecasting 12
2.2 Basic models 17
2.3 Models of global and local validity 19
2.4 Forecasting models and formulae 21
Chapter 3. Forecasting criteria 23

3.1 The need for criteria 23
3.2 Classification of criteria for testing forecasting formulae
on data 25
3.3 Statistical features 29
3.4 Transient features 32
3.5 Steady-state features 32
3.6 Criteria for testing forecasting formulae on models ... 33
3.7 Study by simulation 36
PART II FORECASTING FOR SOME BASIC MODELS
Chapter 4. The constant mean model 41

4.1 The global constant mean model 41
4.2 The local constant mean model 48
IX
X
Chapter 5. Linear trend models 58

5.1 The global linear trend model 58
5.2 The local linear trend model 63
5.3 Polynomial models 75
Chapter 6. Regression models 77

6.1 Introduction 77
6.2 Fitting regression models 79
6.3 Selecting the variables 83
6.4 Econometric models 88
6.5 Time as a regressor variable 90
Chapter 7. Stochastic models 93

7.1 Introduction 93
7.2 Forecasting and conditional expectation . . . . . . .100
7.3 Moving average processes 101
7.4 Autoregressive processes 105
7.5 Autoregressive — moving average models 107
7.6 Models involving stochastic and deterministic elements .111
Chapter 8. Seasonal models 115

8.1 Introductory concepts 115
8.2 Seasonal index methods 124
8.3 Fourier methods 139
8.4 Stochastic methods 147
8.5 Comparison of methods 148
Chapter 9. Growth curves 150

9.1 Models for growth curves 150
9.2 Fitting growth curves 156
9.3 Forecasting growth curves 160
Chapter 10. Probabilistic models 162

10.1 Introduction 162
10.2 Forecasting probabilities 162
10.3 State models 166
10.4 Forecasting criteria 170
10.5 Probability forecasting and decision making 172
Chapter 11. Multivariate models 174

11.2 Generalizing exponential smoothing 175
11.3 Multivariate stochastic models 176
XI
Chapter 12. Forecasting methods and models 179

12.1 Forecasting methods 179
12.2 Model building and identification 179
12.3 Model testing 183
12.4 The use of judgement 188
PART III THE FORECASTING PROCESS
Chapter 13. Data 193

13.2 Sources of data 193
13.3 The quality of data 195
13.4 The adjustment of data 196
Chapter 14. Adaptive methods and other extensions 203

14.2 Adaptive methods 203
14.3 Extensions to recurrence methods 206
14.4 Extensions to the error correction method —
Kalman filters 209
14.5 Linear forecasting formulae 213
14.6 Using mixed methods 216
Chapter 15. The analysis and comparison of methods 220

15.2 Forecast analysis 221
15.3 Choosing forecasting parameters 237
15.4 Comparison of methods 240
15.5 Prediction intervals 241
15.6 Sensitivity analysis 243
Chapter 16. Forecast control 246

16.1 Quality control in forecasting 246
16.2 Tracking signals for bias 248
16.3 Cumulative sum methods 252
16.4 A tracking signal for autocorrelation 254
Chapter 17. Two-stage forecasting 256

17.2 The use of forecast errors 258
17.3 The use of external forecasting information 264
17.4 Combining forecasts 266
17.5 The use of cost and other criteria 268
Xll
Chapter 18. Problems in practice 274

18.2 Practical forecasting criteria — a stock control example . 275
18.3 The forecast as part of a system — a stockholding
example 278
18.4 Forecasts and crystal balls — a scheduling example . . . 281
18.5 The forecast as part of a system — self validating and
self-defeating forecasts 287
18.6 Practical forecasting — a postscript . . .289
Appendix A Some terminology and definitions for stochastic

processes 293
Appendix B Proof that conditional expectation gives minimum

MSE forecasts 298
Appendix C Two-way exponential smoothing 301
Appendix D A technical note on discounted least squares . . . .302
Index 305
Notation
xt an observation at time t
H\ HI H>
a forecast, or estimate, of xt
h a forecast made at time t of xt+h, i.e. with lead time h
the one-step-ahead forecast abbreviated from xtl
et the forecast error xt — xt, often xt — xt_ 1
et h the lead time h forecast error xt+h — xt+h
Mr the underlying mean at time t, i.e. E(xt)
Pt the underlying trend at time t
.
Part I
Preliminaries
Chapter 1
An introduction to forecasting
1.1 Everybody forecasts

Every time we make an appointment we are making a forecast about
our future ability to keep that appointment. At the ‘see you in ten
minutes’ level we do not even think about the forecast, though for
more long-term appointments we try to think of any obvious things
that might stop the meeting from occurring as forecast. The more we
think about our forecast, the more likely it is to be right, for we will
not be so likely to overlook some prior engagement or some other
aspect of the situation. In this example we might even try to assess our
confidence in our forecast by the amount of surprise we would show
when the arrangement falls through. Such acts of unconscious or
semiconscious forecasting are part of our everyday lives; they are also
part of the lives of those who are involved with industrial or business
problems. The good manager is not so much one who can minimize the
effects of past mistakes but rather one who can successfully manage the
future. Consider, for example, the following list of questions:
What will next month’s sales be?

How much should be produced this month?
How big a stock should be kept?
How much material should be bought?
When should the material be bought?
What should the sales target be?
Should the labour force be increased?
What price should be charged?
How many operators are required?
What will the profit be?
To obtain the best anwers to these questions, we would need to be able

to see into the future. To be able to give the best practical answers to
these questions, we need to be able to forecast the future. Our
3
4
forecasts may be wrong, but they must be made. In the past most
answers to questions of the above type have been based on unconscious
or semiconscious forecasts. In these forecasts it was frequently assumed
that the future would be just like the recent past. The growing
interest in many aspects of forecasting is based on the belief that
conscious and careful thought cannot fail to help improve our forecasting
skill and thus our ability to get answers to the above types of question.
It may be in some cases that a thorough scientific forecasting study
might not lead to much better forecasts than the old ‘by guess and by
gosh’ methods. Even so, such a study will enable those involved to have
a better understanding of their situation and thus improve their control
of it
There are many methods of statistical forecasting, none of which can
be put forward as providing all the answers to everybody’s forecasting
problems. The techniques discussed in this book are chosen because
they have found the widest application. The emphasis of the book,
however, is not on specific techniques but on general methods and
attitudes towards forecasting. In this way it is hoped that the reader
will be better able to apply statistical forecasting methods to his unique
problems. New techniques and applications of forecasting are being
published every month, and thus if this book is regarded simply as a
compendium of forecasting techniques it would very soon become
obsolete. If, however, the reader treats it as a text on general methods
and attitudes in statistical forecasting, then he should be able to
incorporate the new developments in this rapidly expanding field into
his basic framework of knowledge.
1.2 Methods of forecasting

In this section we shall define three general methods of forecasting that
involve very different approaches. It is not intended to imply that
these are the only types available, or that any specific method must
automatically be classifiable as one of these, but these are introduced to
provide a general background.
1.2.1 Intuitive methods

These are the classical methods of forecasting. They are based
essentially on the individual’s feeling for the situation, or on the
expert’s opinion. Often one is forced into these methods by the
complexity of the situation and the sparceness of reliable and relevant
data. One obvious way of seeking to improve these methods is to
replace the single expert by a small group or larger number of people.
For example, in forecasting sales one might use the summarized views
of the sales force or of a committee of the area sales managers.
5
Alternatively, a sample survey of the views of potential customers

might yield more reliable information.
1.2.2 Causal methods

These methods are based on trying to forecast effects on the basis of
knowledge of their causes. In many forecasting situations the causes are
economic and hence most of the work in this field is also of an
economic nature. Economic relationships between causes and effects
are studied; as the effects occur after the causes direct knowledge of the
causes leads to a forecast of the effects. But if, as often occurs, the time
lag between cause and effect is very small, we may still have to forecast
the causes, since information on the causes may not become available
until some time after the effects have occurred. The causes used for
forecasting are then chosen as those that can be forecast in a more
reliable fashion than the effects. If the mechanism of the relationship is
known, quite elaborate forecasting methods may be devised. Often,
however, the nature of the relationship is obscure and we are forced to
use some purely empirical form of relationship. Chapter 6 gives a
discussion of some of the relevant forecasting methods. There are a
considerable number of books dealing with economic forecasting; the
references list some examples.
1.2.3 Extrapolative methods

These methods are all based on the extrapolation into the future of
features shown by relevant data in the past. The methods are usually of
a mathematial or statistical nature and provide the basis for most of this
book. Such methods are sometimes referred to in books on economic
forecasting as naive methods. However, to understand any situation we
must start by being naive, and consider the simplest methods and
approaches. When we understand these thoroughly, we may then progress
to the higher realms.
1.3 The basis of forecasting

To attempt a scientific study of forecasting the future might seem a
hopeless exercise, since each problem is obviously unique. As we look
more closely at specific examples of forecasting, however, we find some
common priniciples. Let us consider some examples.
Example 1.1
Most political policy-makers are concerned with forecasting the
future as it would be if no new actions were taken. They then consider
6
the structure of society as they see it and take action to introduce

policies that will improve society according to their political criteria. To
be able to decide what action to take, they must be able to identify
structures in society that will respond to their policies. It is not
unknown for the implemented policies to fail. This might be due to an
incorrect identification of the structure or, alternatively, to a lack of
stability in that structure. To correctly make such forecasts and define
policy, the policy-maker needs to be able to identify the appropriate
structure and have some confidence in its stability over the
appropriate period.
Example 1.2
H. T. Davis (1941) points out that had the mass of the sun not been
so great, relative to the planets, statistical methods would have had to
be used to investigate the physical laws of gravitation and to predict
future movements of the planets. As things are, the gravitational pull of
the sun on a planet is so powerful that the planet’s orbit is almost the
same as it would be if the other planets did not exist. From
observations on this orbit the structure of the relationship between sun
and planet can be inferred and the inverse square law of gravitation
deduced as the model that fits most of what is observed. The great
stability of the situation enables very accurate forecasts of the future
positions of the planets to be made. Had the mass of the sun been
smaller, its effects would not have dominated those of the other planets
and much more complicated orbits would have occurred. The conse-
quence of this complication would have been that, though the inverse
square law would still be the basic law, it would have been much harder
to find it from the observed data. Further, forecasting the future
positions of the planets would have been much more difficult, as the
simplicity and stability of orbit we have would not have occurred.
Example 1.3
As a third example we will consider some of the forms taken by a set of

data which might correspond to orders received by a firm each week.
Table 1.1 gives the appropriate data. Suppose the form of the data is
that given in line (a) of the table. Here the orders come from just one
customer whose demand is constant. The data show a very simple
structure, constant demand, and a perfectly stable behaviour. We would
thus feel fairly confident in forecasting orders of 6,000 items for week
13. Suppose, however, the data took the form shown in line (b) of
Table 1.1. Here the occurrence of 9,000 units of demand at week 5
leads to some doubt about the stability of the structure shown in the
remaining weeks. Hence we might still forecast 6,000 items for the next
7
Table 1.1 Orders received (in 000s)
Week number
1 2 3 4 5 6 7 8 9 10 11 12
(a) 6 6 6 6 6 6 6 6 6 6 6 6
(b) 6 6 6 6 9 6 6
(c) 6 6 6 6 9 6 6 6 6 9 6 6
(d) 7 6 6 4 7 6 8 6 5 6 7 5
(e) 8 8 9 8 12 12 15 14 14 16 18 17
week, week 8, but no longer with such confidence. Suppose that we

wait a few more weeks to see what happens and enquire into the reason
for the demand for 9,000 units. We may find that the data in (b) arise
from orders by two customers, the one referred to above and also
another customer who orders 3,000 items at regular five-weekly
periods. The sensible forecast now would be 6,000 items for weeks 13
and 14 and 9,000 items for week 15. We see then that it is sometimes
possible to cope with what appears to be an instability in a structure by
considering more sophisticated forms of structure. The apparent insta-
bility of data (b) becomes the stable structure when more information
is obtained, as in data (c).
In practice, data such as those in (a), (b) and (c) do not occur very
frequently. We are much more likely to see data such as those shown in
line (d) of Table 1.1. These data do not show the nice simple
mathematical structure of (a), but nonetheless they do show a
structure, as is seen from Figure 1.1. This gives a simple bar chart of the
data, showing the frequency with which the sizes of orders occur, and it
reveals a quite clear structure, but in this case it is a statistical structure
rather than a mathematical one. It is apparent that the average order is
near to 6,000 units, and, on the evidence of this small sample of data,
the orders appear to be distributed fairly evenly about this value with
decreasing frequencies of occurrence away from the average.
Frequency 5
’| _L
4 5 6 7 8
Orders (in OOOs)
Figure 1.1 Bar chart of the data of

Table 1.1(d)
8
The structure of this situation is thus a structure in the probabilities

of orders taking on certain values. To describe such a structure we will
need the language of statistics as well as mathematics. When the term
‘stability’ is applied to this situation, it must be applied to the whole
statistical structure. The values of the orders are not stable in the
mathematical sense, as they vary from week to week in a chance
fashion. What is assumed to be stable is the structure of this chance
behaviour as illustrated for the sample of 12 weeks by Figure 1.1. In
forecasting the orders for week 13 it would be reasonable to forecast
6.000 units, but now we are not completely certain of our forecast as
we were for data (a) and we are not subject to vague doubts as we were
for data (b). The information in Figure 1.1, i.e. the statistical structure,
enables a much clearer idea to be formed of the possible errors in the
forecast. If we take the data as providing estimates of the probabilities
of error, and consider only units of thousands, we have a probability of
about S/12 of being right with our forecast of 6,000, about % 2 of being
1.000 units in error and l/6 of being 2,000 units in error.
Figure 1.2 Plot of the data of Table 1.1(e)
Consider, finally, the set of data in line (e) of Table 1.1 which is
plotted in Figure 1.2. It is clear from the graph that the data lie in a
‘chance’ fashion about a line. This might occur if, added to the source
of the orders in data (d), and additional customer ordered 1,000 units
in week 1, 2,000 units in week 2, and so on. Thus in mathematical
terms his orders in week n are 1,000 n. It is thus clear that the structure
of the data has both a mathematical and a statistical aspect. To say that
such a structure is stable, we imply that both the upward trend and the
statistical variation about that trend are both stable. The forecast for
9
week 13 would be our forecast of 6,000 units as made for week 13 in

section (d) plus the 13,000 units from the trend, the new customer,
giving 19,000 units. The knowledge of the statistical structure would
enable exactly the same probability statements to be made as were
made for the forecast for data (d).
Example 1.4
Table 1.2 gives two years’ data and Figure 1.3 gives a plot of this set of
data. At first sight these data look unstructured. A study of the situation
in which it was obtained suggests that it might be composed of an
upward trend, some sort of seasonal variation and a chance variation, or
as it is usually called, a random variation. Table 1.3 shows a breakdown
of the data into these three elements, together with a table of the
frequencies of the random variation, which gives the same sort of
information as Figure 1.1. Thus the data have a structure composed
of both mathematical and statistical elements. It is clearly seen from
Table 1.2 Two years’ data
Month
Year 1 2 3 4 5 6 7 8 9 10 11 12
1 33 34 30 36 26 22 14 25 19 20 38 37
2 42 49 42 33 37 34 23 40 23 31 44 40
Figure 1.3 Plot of the data of Table 1.2

10 \
Table 1.3 Breakdown of data in Table 1.2

(a)
Random
Trend Seasonal variation Total
Month Year 1 Year 2 additions Year 1 Year 2 Year 1 Year 2
1 21 33 10 2 -1 33 42
2 22 34 12 0 3 34 49
3 23 35 10 -3 -3 30 42
4 24 36 8 -1 -11 31 33
5 25 37 0 1 0 26 37
6 26 38 -6 2 2 22 34
7 27 39 -16 3 0 14 23
8 28 40 -10 7 10 25 40
9 29 41 -9 -1 -9 19 23
10 30 42 -7 -3 -4 20 31
11 31 43 2 5 -1 38 44
12 32 44 6 -1 -10 37 40
(b) Frequency table for random variation
Values Less than —7, --7 to —5, —4 to —2, —1 to 1, 2 to 4, 5 to 7, 8 to 10

Frequency 3 0 4 9 5 2 1
this example that finding the underlying structure for a set of data can
be a difficult task. It is certainly not obvious that the data of Table 1.2
were obtained in the fashion shown.
It is also clear that, given enough effort and ingenuity, one could
create some such structure for any set of data. If, however,we are going
to assume stability of the structure for forecasting purposes, it is clear
that the structure that is used must correspond to some real structure in
the situation. The mathematical and statistical equations used in
forecasting must make sense in terms of the situation in which they are
used. The search for a structure for use in forecasting is thus not simply a
mathematical or statistical task but requires a broad knowledge of the
nature and background of the particular forecasting problem being
investigated.
We have seen from the above examples that the basic requirement for
forecasting is the existence of a stable structure. This structure may be
of purely mathematical form or purely statistical form or, most
frequently, a mixture of both. So far these structures have been dealt
with in a purely descriptive fashion; to proceed further we must be
more precise and rigorous in the description of the structures.
Mathematical and statistical models of forecasting situations provide
such a rigorous description. It is an introduction to the study of such
models that provides the subject matter for the next chapter.
11
References
Chisholrne, R. K., and Whitaker, G. R. (1971). Forecasting Methods. R. D. Irwin,
Homewood, Illinois.
Davis, H. T. (1941). Analysis of Economic Time Series. Principia Press.
Morrell, J. (1972). Management Decisions and the Role of Forecasting. Penguin,
Harmonds worth.
Robinson, C. (1971). Business Forecasting. An Economic Approach. Thomas
Nelson and Sons, London.
s
Chapter 2
Models
2.1 Scientific forecasting
It has been seen in Chapter 1 that forecasting is an inherent part of

business and industrial activity and that until recently it has been done
mainly by intuitive methods. An essential part of the present revolution
in business methods is the attempt to use scientific methods as an aid to
making business decisions. Let us consider now some of the steps which
are involved in applying scientific method to forecasting problems.
2.1.1 The first step — data collection
Before we can apply any scientific forecasting method to a situation,

we must naturally examine the situation in great detail. In particular,
this involves gathering together as much relevant data on the situation
as possible. Such data are of two main types: firstly, data which are
internal to the firm, e.g. records of past sales, production, stocks, etc.;
secondly, data which are external to the firm. For example, Govern-
ment and Trade Association statistics often provide a background
against which to interpret the internal information. Sometimes this
external information may be of direct value in producing a forecast. An
example of this would be when forecasts for a whole industry are used
to provide basic information for forecasting the sales of a firm within
that industry.
2.1.2 The second step — data reduction
The aim here is to pick out from the possibly vast amount of
information obtained at the first step that which is regarded as most
relevant. This is then reduced to a bare minimum. For example, in a
sales forecasting situation the record of past sales for the firm would
probably provide this minimum information. In deciding which items
12
13
are to be included in this basic minimum information three criteria are

frequently used. The minimum information should meet the require-
ments of
Relevance: Is the information the most directly relevant information

available?
Reliability: How were the data obtained? How trustworthy are
they?
Recency: Is it the latest information available? Will new information
be available promptly?
When this minimum information has been thoroughly studied and its
behaviour understood, we might start the reverse process of allowing
other information to be used. For example, in sales forecasting we
might use economic information or the assessments of the situation
made by the sales force. However, for the moment we will consider
only the use made of this minimum information. Though we have
produced some criteria to be used in the selection of our information,
we cannot usually do this in a purely objective fashion. Selection must
be made on the basis of a broad and deep understanding of the practical
situation. Later scientific study of this information will indicate
whether our choice was good or bad. This will reveal itself by our
success in obtaining good forecasts from the information. Unfortun-
ately, we may not find out until too late that our information is not the
best, and so our intuitive choice is a necessary and vital step in the
process.
2.1.3 The third step — model construction
We now examine our minimum information with the object of

identifying its structure and producing a precise specification for this
structure. In most cases this specification will consist of a mathematical
or statistical description of the structure; this description is what we
normally call the model for the structure. To make this idea more
concrete, let us consider again the data given in Table 1.1. In the
previous chapter we sought to describe these data in purely verbal
terms. Now we will describe them in mathematical terms. Let xt denote
the orders in week number t, t = 1, 2, 3, . . . . For data (a) xt is always
6,000 units, so we write
xt = 6,000 (t = 1, 2, 3, . . .)
For data (c) xt becomes 9,000 in weeks 5, 10, 15, . . . , and so we write
x, = 6,000 {t = 1, 2, 3, 4, 5, 6, 7, 8, 9, 11, . . .)
xf = 9,000 (t = 5, 10, 15, 20, . . .)
For data (d) we found that the orders differed from 6,000 by a value
14
0 4
02 0 2
0 1 |
°'l
-2-10 I 2
Difference in orders from 6,000 (in 000s)
Figure 2.1 Probability distribu-

tion
that was not determinate but was governed by a statistical structure.

Let us use the symbol et to denote the chance or random difference
from 6,000 that occurred in week t. This quantity et can take on values
. . . , —2, —1, 0, 1, 2, . . . (000s) with certain probabilities. The data
given in Figure 1.1 gave the results of taking a sample of these random
differences. Suppose the values given in Figure 2,1 gave the actual
underlying probabilities of et. The structure of the situation can thus be
imagined as a vast hat containing tickets labelled —2,000, —1,000, 0,
1,000 and 2,000, in the proportions given in Figure 2.1, these
proportions in fact being the probabilities of choosing the respective
ticket numbers. At the end of week t a ticket is chosen at random, its
number we call ef, and this number is added to 6,000 to give xt. Thus
the model is
xt = 6,000 + et
where, to use more precise terms, et is a random variable having the

probability distribution given in Figure 2.1. For the data of line (e) of
Table 1.1 a trend is added to the structure. This trend, as shown in
Figure 1.2, is expressed mathematically as
Trend value at week t = 1,000/
Hence the orders are modelled by the equations
xt = 6,000 + 1,000/ + et
In our example we have taken the data as the starting point and then
constructed the model. The illustration of the tickets in the hat suggests
that it is always possible to start with a model and construct from it
data having that structure. The purely mathematical part of the model
presents little problem, but generating values for et can get quite
complicated. A literal hat would do the job for the distribution in
Figure 2.1. For more complicated distributions we need more elaborate
equipment, most usefully a computer. There are two reasons why it is
useful to be able to generate forged data having the properties of any
given model. The first is that it gives us experience of the behaviour to
15
expect from data of the given structure, and, secondly, it enables us to

examine the properties of different methods of forecasting. The
experience gained of the behaviour of different kinds of model is a
great help in deciding what model to consider for our own particular set
of data.
If we used the probability distribution of Figure 2.1 to generate a set
of et to use in the trend model, we would obtain a set of data which
was not identical to our original data in Table 1.1 but which would
have the same structure. Thus we see that the model is not identical
with the set of data, nor does it completely explain why the data
behave in this or that detailed fashion. The model simply gives a
mathematical or statistical formulation of the situation, a formula that
can generate numbers showing a similar behaviour to the observed data.
2.1.4 The fourth step — model extrapolation

The important difference between the model and the data for our
purposes is that the latest piece of information completes the data,
whereas the model will produce values for any value of t we care to
substitute. Thus in the expression
xt = 6,000 + 1,000£ + et
we can substitute any value for t and not just the values t = 1 to 12
which represented the original data. The evaluation of the mathematical
part, 6,000 + l,000f, for differing values of Ms a matter of simple
substitution. The evaluation of the random part, et, will depend on
what we are seeking to do. If the aim is to generate a set of data with
the given properties, a numerical value must be found for et from the
appropriate probability distribution. If the aim is to forecast, a
reasonable procedure would be to replace the unknown future value of
et by its theoretical average value, which is usually zero. Thus a forecast
of Xi 4 , denoted by ocl 4, would be
x! 4 = 6,000 + 1,000 x 14 + 0
= 20,000
since in the distribution of Figure 2.1 the theoretical average, or
expectation, of et is zero. We thus extrapolate our model, our structure,
into the future, assuming that the situation is stable.
The basic steps that we have discussed are represented in Figure 2.2.
There are many problems associated with the construction of models
that have been ignored in the previous general discussion, in particular
the problems of choosing the form of model and of choosing the
numbers to use in it. For example, how did we know that in the above
illustration a trend model should be used with a starting value of 6,000
and a constant rate of increase of 1,000 units per week. In a real
16
Figure 2.2 Four basic steps in statistical forecasting
situation we start with complete ignorance, i.e. with the model

xt = ?
To decide what this question mark should be is called the problem of

IDENTIFICATION, the problem of the choice of model. In the above
data we have assumed that we have solved this problem and we have a
model of the form
xt = a +13t + et
where a and (3 are constants, /3 being the slope of the line and a the
intercept with the x-axis. We still, however, do not know what
numerical values to put for the constant a and the slope (3. The problem
faced here is called the problem of ESTIMATION or model fitting, the
choice of numerical values to use in the model. When we have solved
this problem for the above model, we then might put
xt = 6,000 + 1,000£ + et
Notice that these are not once and for all problems to be dealt with
before we start forecasting. At any stage our model will, we hope, be
one that gives a good description of the data so far observed. However,
as time progresses new data will be obtained and the structure may
change or the increased amount of available information may reveal
features that were not apparent before. There are two consequences of
this possibility. The first is that the processes of identification and
estimation need, in some fashion, to be continued as long as
forecasting is continued. The second is that even if to date we have a
very good model for the data which has given excellent forecasts, there
17
is still no guarantee that this model will continue to give good forecasts
in the future. The chances of a structure remaining stable decrease, the
longer the period of time considered. It follows that the doubt about
the reliability of our forecasts increases, the greater the period into the
future over which we extrapolate our model. The possible failure of our
basic underlying assumption that we have correctly identified a stable
structure means that the forecaster can never quite throw out his
crystal ball.
2.2 Basic models

It was seen in our previous discussion that the choice of model is a basic
step in forecasting. A natural reaction to this by many practical people
is to feel that since their situation is probably unique they will have to
invent their own models to fit their situation. Since they may feel this
to be beyond them, they might lose hope in statistical forecasting.
Fortunately, it has been found in practice that a very wide range of
data can be adequately modelled by a relatively small number of basic
models. Table 2.1 gives a list of some of these basic models. Part II of
this book considers the use of each of these models in forecasting.
Often these models provide all that is needed for practical forecasting.
Sometimes, however, they will need to be combined, modified and
tailored in some fashion to the problem in hand. Some approaches to
doing this and some examples are given in Part III of this book.
Let us use the models of Table 2.1 to give a first introduction to
some terminology. Models such as (a), (b) and (c) in the table are
perfect mathematical forms. For example, if we had at times 0, 1, 2, 3,
4, data 2, 4, 6, 8, 10 from the perfect linear trend, then any two of the
observations could be used to show that a = 2 and /3 = 2; so a perfect
forecast of the next observation would be 12(= 2 + 2 x 5). Such a
model we will term a deterministic model. As we discovered in example
1.4, such models rarely describe the real world owing to the chance, or
random, variation that is shown by most data. Random sequences such
as that shown in (g) of Table 2.1 are often used to model this aspect.
So, as we have seen, we might have data modelled by
xt = linear trend + random sequence
or, mathematically,
xt = a + fit + et
where et is used to denote a sequence of independent random
quantities scattered above and below zero. As the deterministic part is a
fixed basis for this model, we will still often refer to this type of model
as essentially deterministic. Unfortunately, data from such a model will
look like that shown in Figure 1.2, so it will not be possible to take just
18
Table 2.1 Some common models

.— ■■■■■■ H . MM • I ■■■■■■■ ■■■■■»■■— ■ ■
Name Plot Form
(a) Constant xt = ix for all t

mean
(b) Linear trend xt = a + fit for all t
(c) Linear xt = a + fizt

regression zt being another variable
(d) Autoregression xt = a + fixt_l

xt_ i being the previous value
of x
(e) Periodic xt repeats pattern every T

observations
=
Xf+nT f®^ t
(n = 1,2,3)
(f) Exponential xt = a e^f for all t

growth
(g) Random normal xt is normally distributed

sequence with mean ix and variance a2
values of xt independent of
each other for all t
19
two observations to find a and (3. We will need to develop methods to

find values for a and (3, to estimate a and (3, so that the model we get
shows a good fit to the data and thus can reasonably be extrapolated to
obtain a forecast. The problems of estimating parameters, such as a and
f3, and thus fitting models to data will be basic to much of our
discussion in the following chapters.
Let us look now at the model described by (d) of Table 2.1. As it
stands, the situation described by this equation is deterministic.
However, if we become more realistic and write
X, = a + Pxt_ ! + e,
to allow for the chance variation of the real world, we obtain a very
different situation. As xt is influenced by the previous x, it must also be
influenced by previous random variables e. Thus xt is influenced by the
whole past history of the sequence of es. Such a model is an example of
a stochastic model. Such models form the base for a large number of
forecasting methods. Many forecasting methods are based on either a
deterministic or a stochastic model. We will tend to look at them
separately, but clearly there are situations where one would seek to
combine these types of model.
Models (b), (c) and (d) all have the form
xt = a + /3( )
The parameters thus occur singly in separate terms of the model, the
terms being added together. The influence of the parameters and the
effects of changing them can thus be considered independently for each
parameter. Such models are said to be linear in the parameters.
Problems of estimating parameters and forecasting tend to be simplest
for linear models. Other models, such as the growth curve in (f) in
Table 2.1, do not have this property. Here the model is non-linear in
the parameters; the effect of changing (3 in (f) will depend on the value
taken by a. In studying such growth curves we will be faced with more
difficult problems of estimation and forecasting.
2.3 Models of global and local validity

When we use models such as those given in our list, we shall sometimes
find that they are being used in a much more rough and ready fashion
than is indicated by the mathematics. In particular, we shall often use
them not in the belief that they exactly describe the underlying
structure of the situation but in the faith that, at least in the recent past
and, one hopes, the near future, they will give a reasonable description
of this underlying situation. We shall thus consider two possible
situations that could occur in practice. In the first, the structure is
regarded as highly stable and the chosen model as the truth about the
20
underlying structure of the data. The model for such a situation will be
called a global model. In the second situation, the structure is believed
to be stable in the short run but not necessarily in the long run. Slow
changes in structure may occur in our data which will not affect us too
adversely for forecasting over short periods of time. When examining
forecasting at an advanced level, attempts are usually made to
incorporate any slow changes in structure into a more complex model.
It is, however, often sufficient to treat such data as coming from, what
we will term, a model of local validity, or a local model.
There is no difference in the mathematical or statistical formulation
of these two types of model, global and local. The difference lies in the
way in which we make use of the models. Thus when we use a local
constant mean model we shall forecast the future on the assumption
that the mean in the near future will be the same as the current mean.
We will not be willing, however, to assume that the current mean is the
same as that in the past when we first obtained our data. We will, in
future chapters, discuss models from two viewpoints: firstly, from the
global viewpoint, with the model assumed to be a proper description of
the underlying situation; secondly, from a local viewpoint, with the
model regarded as an approximation valid only locally in time. There
are a number of different situations that would lead to the use of
models as approximations only. These are
(a) situations where there is not enough data available to enable us to

study the underlying situation in detail;
(b) situations where for reasons of simplicity and economy we
deliberately use a simpler model than the one that we think might
be appropriate;
(c) situations where the quantity being forecast is known to be
influenced by factors that are not stable but which change in time
in a relatively slow fashion.
Experience suggests that using models as local approximations often

leads to better forecasts than are obtained from their global use. That
this is so may well relate to the fact that with data in time we are
always effectively in situation (a). The underlying situation by its
nature stretches from some past time to some unknown future time.
Even if by magic we had data from all this time, it would only represent
one observation on the entire process. In practice we only have data
from a small section of time and thus, in effect, we have less than one
observation. That we can get anything at all from this situation depends
on major assumptions about the nature and stability of the process.
Against such a background, using our models as approximations to
sections of the underlying process seems a safer course than using our
models as global descriptions of the situation.
21
2.4 Forecasting models and formulae
People who have studied statistics will be familiar with the ideas of
statistical models for data, such as those just referred to. It is therefore
important to draw a distinction between the use of models in
forecasting and the more usual statistical uses. In fact, we might refer to
the models in section 2.3 as forecasting models. Though they are
statistical or mathematical models in form, their required relation to
the data is different in forecasting from their relation in ordinary
statistics. In statistical model building we are just as interested in how
the model describes the first observations as in how it describes the
most recent observations. In forecast model building it is the future
that interests us, and as long as the model is good at giving forecasts we
do not mind how badly it fits the data of the dim and distant past.
Thus we will often feel justified in using a model that obviously does
not fit the entire set of data. Provided it gives a good fit to the most
recent data, it is a potentially good forecasting model. Putting this a
different way, our attitude to models in forecasting means that we are
as happy with a local model as with a model of global validity, as long
as we get good forecasts.
Much of this book will be devoted to describing how to obtain
formulae for forecasting future values in particular forecasting models.
We will call such formulae forecasting formulae. For example, given a
model of the form
xt = n + et
where p is a constant mean and et is a purely random normal sequence,
and given past data xx, x2, . . . , xf, we wish to forecast the future value
xt+k. Denote by xtk the forecast made at time t of the value of x at k
time units into the future, referred to as a forecast with lead time k. An
appropriate forecasting formula for xt+k is
=
%t,k ~ (xx + x2 +...+#*)
Usually the forecasting formulae will be obtained from some specific

forecasting model. However, it often occurs that a particular forecasting
formula provides reasonable practical forecasts for a number of
forecasting models. Hence we will sometimes be interested in fore-
casting formulae in themselves without reference to forecasting models.
Another forecasting formula that might have been used in our
example, had we felt that the model should be used as a local model
rather than a global model, is
xt, k ~ (1 a xta xt_ 2 • • •)

where a is some number between 0 and 1 which is chosen to give the
22
best forecast in some sense. This number is a parameter of the

forecasting formula and will be referred to as a forecasting parameter,
as distinct from a model parameter such as ju. The choice of such a
forecasting parameter will control the properties of the forecasts. Thus,
if in the example a is small, the formula will strongly emphasize the
recent values, xt, xt_x , etc. If, however, a is chosen close to 1 and t is
large, the formula gives much the same answer as that given by the first
formula.
Chapter 3
Forecasting criteria
3.1 The need for criteria

Consider the following four problems:
Problem A
A sales manager collects together records of past monthly sales for a

particular product. How does he now find a forecasting formula that
gives good forecasts for this data?
Problem B
A member of his staff suggests a particular method of forecasting.

How does the sales manager test this method to see if it is any good?
Problem C
Having examined several products, the sales manager decides that

they all have a similar structure which could reasonably be modelled by
a linear trend with an added random variation adequately described by a
purely random normal sequence. How does he now find a forecasting
formula that gives good forecasts for this particular model?
Problem D
His enthusiastic member of staff advocates his pet method as the best
method for this type of model. How does the sales manager test the
success of this method for his model?
The ability to find answers to questions such as these is obviously

basic to the whole exercise of practical forecasting. However, in each
problem it is assumed that we know what is meant by a good forecast.
23
24
We forecast sales of 6,539 items arid, lo and behold, we sell 6,539

items — a perfect forecast! But suppose we sell 6,502 items; can we say
that our forecast was good or not? Without more information we
cannot answer this question. The error of 37 items may be large or
small relative to an acceptable error in the given situation. We cannot
say whether a forecast is good until we are agreed on what is meant by
the term ‘a good forecast’. The two words ‘good’ and ‘forecast’ are
really inseparable A bad forecasting method is just not a tenable
method of forecasting. The consequence of this situation is that we
must postpone a discussion of how to forecast until the nature of a
‘good forecast’ has been investigated. The aim of this chapter is to carry
out this investigation.
A careful examination of the four problems at the beginning of this
section shows that the subject matter to which the criterion of
goodness is applied may be different in different situations. In two of
the problems the subject matter is the data itself; in the other two it is a
theoretical model. Further, the objectives may be different in different
situations. In two of the problems the object is to derive a forecasting
formula; in the other two it is to test out a given formula. There are
thus four classes of problem,as is shown in Table 3.1.
Table 3.1 A classification of problems needing forecasting criteria
Objective of use of criteria

Subject matter
of criteria Deriving a forecasting Testing a forecasting
formula formula
Data A B
Model C D
We found in Chapter 2 that obtaining a forecasting formula, given a

set of data, involved first replacing the data by some appropriate model.
Thus, obtaining a formula for a set of data usually reduces to the
derivation of a formula for a model. The word ‘usually’ is inserted here
since we will later look at methods where the model-building aspect is
ignored. The ‘best’ formula is often found by testing out several
forecasting methods using the data and choosing the best according to
some criterion. In this case the problem of deriving a formula is
replaced by the problem of testing a formula. Thus, in general,
problem A is reduced, most frequently, to problem C, though some-
times to problem D.
The problem of deriving formulae for a given model to meet certain
criteria can be a very difficult one. It will be dealt with at different
25
levels at various parts of this book It is sufficient to note at this stage

that, as is reasonable, the criteria that can be used as the basis for
deriving a forecasting formula for a given model are usually the same as
those which are used for testing out any given formula on the model.
Exceptions to this rule sometimes occur, in particular a forecasting
formula may be first derived according to one particular criteria and
afterwards tested out against other criteria to provide additional
information about its behaviour. However, for the purposes of this
introductory chapter, we will put together problems of type C with
those of type D under the latter heading.
The outcome of the above discussion is that a preliminary investi-
gation of forecasting criteria may be made by considering only the case
of the problem of testing a given forecasting formula. We consider in
particular how to measure the success of a given forecasting formula (a)
for a particular set of data and (b) for a particular model.
3.2 Classification of criteria for testing forecasting formulae on data

In this section we shall consider the situation where we have a set of
data, ^ X ^
iiX/ | 2 ^ ^ ^ and a forecasting formula which can be used to
obtain a forecast from such data. We will denote the forecast of xt
by xt9 where xt, is obtained from the forecasting formula using some or
all of the observations that existed before time t. The discussion of
forecasting criteria is thus a discussion of the sense in which xt might be
close to xt.
To make the discussion as concrete as possible, consider the data in
Tables 3.2 and 3.3. It will be seen from Figure 3.1, which charts the
data of Table 3.2, that the first part of the data is constant at 6 units,
the second section shows random variation about this constant value
and then, after the fifteenth observation, the underlying mean of the
set of data jumps from 6 units to about 10 units. The behaviour of the
data in Table 3.3 is similar to that of Table 3.2 for the first six
observations, but then there is a steady upward trend. Columns xA and
xB of these tables and the corresponding plots in Figures 3.1 and 3.2
show the forecasts for this data obtained by two different forecasting
formulae, A and B. It is seen that both methods give perfect forecasts
where the data is constant at the value of 6. Where the data show
random variations, both forecasts also show random variations; how-
ever, the variability in the forecasts obtained by method A is much
greater than that shown by the method B forecasts. Method B is said to
‘smooth’ the data more than method A. Where there is a step change in
the underlying mean for the data, method A very quickly picks up the
new mean and varies randomly about the new value. Method B is very
slow to follow the data and even after some ten observations the
forecast is still lagging considerably below the new mean value. Where
26
Table 3.2 Illustrative forecast results I
Time Data Forecasts Errors

t X XA e
A
1 6 — —
2 6 6 6 0 0
3 6 6 6 0 0
4 6 6 6 0 0
5 6 6 6 0 0
6 6 6 6 0 0
7 7 6 6 1.00 1.00
8 5 6.5 6.10 -1.50 -1.10
9 4 6 5.99 -2.00 -1.99
10 8 4.5 5.79 3.50 2.21
11 6 6 6.01 0.00 -0.01
12 7 7 6.01 0.00 0.99
13 7 6.5 6.11 0.50 0.89
14 5 7 6.20 -2.00 -1.20
15 10 6 6.08 4.00 3.92
16 9 7.5 6.47 1.50 2.53
17 11 9.5 6.72 1.50 4.28
18 12 10 7.15 2.00 4.85
19 10 11.5 7.64 -1.50 2.36
20 9 11 7.87 -2.00 1.13
21 8 9.5 7.99 -1.50 0.01
22 10 8.5 7.99 1.50 2.01
23 11 9 8.19 2.00 2.81
24 10 10.5 8.47 -0.50 1.53
25 9 10.5 8.62 -1.50 0.38
Figure 3.1 Plots of data and forecasts from Table 3.2

27
Table 3.3 Illustrative forecast results II
Time Data Forecasts Errors Transients

A A
t x -V*
XA
-V*
e
A e
B 2
A Z
B
1 6 — —
2 6 6 6 0 0 0 0
3 6 6 6 0 0 0 0
4 6 6 6 0 0 0 0
5 6 6 6 0 0 0 0
6 6 6 6 0 0 0 0
7 7 6 6 1.0 1.0 0.5 9.00
8 8 6.5 6.10 1.5 1.90 0 8.10
9 9 7.5 6.29 1.5 2.71 0 7.29
10 10 8.5 6.56 1.5 3.44 0 6.66
11 11 9.5 6.90 1.5 4.10 0 5.90
12 12 10.5 7.31 1.5 4.69 0 5.31
13 13 11.5 7.78 1.5 5.22 0 4.88
14 14 12.5 8.30 1.5 5.70 0 4.30
15 15 13.5 8.87 1.5 6.13 0 3.87
16 16 14.5 9.49 1.5 6.51 0 3.49
17 17 15.5 10.14 1.5 6.86 0 3.14
18 18 16.5 10.82 1.5 7.18 0 2.82
19 19 17.5 11.54 1.5 7.46 0 2.54
20 20 18.5 12.29 1.5 7.71 0 2.29
21 21 19.5 13.06 1.5 7.94 0 2.06
22 22 20.5 13.85 1.5 8.15 0 1.85
23 23 21.5 14.67 1.5 8.33 0 1.77
24 24 22.5 15.50 1.5 8.50 0 1.50
25 25 23.5 16.35 1.5 8.65 0 1.35
Large t t t- 1.5 t- 10 1.5 10.00 0 0.00
there is a perfect linear trend, the forecasts obtained by method A lag a

fixed 1.5 units behind the observations. Method B also gives forecasts
that lag behind the data, but it is much slower at ‘realizing’ that a
change has taken place, and even after 20 time units the forecasts are
not increasing at the same rate as the observations, though from
Figure 3.2 it is seen that its rate of increase is slowly catching up with
that of the data. When the rates of increase become the same, the
forecast will lag behind the data by a constant 10 units.
Let us return now to the basic problem of what constitutes a
criterion of success for a forecasting method when applied to a set of
data. From the examples it is seen that there are various ways in which
the forecast can differ from the data. There is thus not just one
criterion of success in forecasting but many. Two of the great
difficulties in practical forecasting are the choice of which criteria to use
and the balancing of opposing criteria to get the best of several worlds.
For example, we might require both that a forecast does not pick up
28
Figure 3.2 Plots of data and forecasts from

Table 3.3
too much of the random variation in the data, as method A does, nor
that it lags too far behind any new feature, as method B does. When
dealing with testing out forecasting formulae, both these aspects must
be studied. It must be recognized, however, that if a method is to be
sensitive enough to adjust quickly to major changes in the structure of
the data, it will also be sensitive to the random variations in the data.
Conversely, if we wish to reduce the response to random variation,
giving forecasts that are ‘smoother’ than the data, we will also prevent
the forecast from responding quickly to changes in the structure of the
data.
It is of value to classify the features that interest us when trying out
forecasting formulae on data (and also on models). Three main headings
can be used. These are as follows.
(a) Statistical features

This heading includes the amount of smoothing produced by the
formula and the statistical distribution of the forecast errors.
(b) Transient features

These are the features that show up when a change occurs in the
underlying structure but which in time die away, e.g. the differing rates
of response to the new trend in Table 3.3 and Figure 3.2.
29
(c) Steady-state features

These are the features shown by forecasts when used in a stable
situation. An example of these is the constant lag shown by both
methods in Table 3.3 after the transient effects have died away.
A detailed study of these features is usually only possible when

testing forecasting methods on models. When considering a set of data,
lack of knowledge of the underlying structure of the data makes
detailed study under these headings difficult and tends to blur the
distinction between them. However, they are still of sufficient use to
justify their separate consideration in more detail.
3.3 Statistical features

To study the elementary statistical features of a situation, it must be
assumed that the situation has a stable structure. For this to be the
case when forecasting the data must have a stable structure and hence
there should be no transient features in the situation. If there are clear
changes in structure in the data, the study of statistical properties
should be based only on data sufficiently far from the changes that the
affects of transient features can be neglected.
To illustrate the statistical features, consider observations 7 to 14 in
Table 3.2. These show a simple random variation about a constant
mean. Table 3.4 repeats this data for ease of discussion. The fourth and
fifth columns of this table show the forecast errors et, which we will
always take as being actual xt minus forecast xt,
et ~
Thus a positive error means that the value that occurs is larger than the
forecast value. It should be noted that some writers define the error the
other way round, as xt — xt. An examination of the errors in Table 3.4
shows that the errors from method A show a greater spread than those
of method B, but the average error is smaller. If we examine the average
error in the data of Table 3.2 for the trending region, this average will
be positive, 1.5 for method A, indicating the lag of the forecast behind
the trend. Hence a study of the errors requires at least a study of
(a) their average (mean) value and (b) their spread. The mean error e for
a set of errors ex, e2 , . . . , en is defined simply as
e=\(ex +e2+...+e„)
1 H
= — Z et (using the summation notation)

n i= i
A good method of forecasting will usually be required to give a small

30
Table 3.4 Statistical analysis of forecasts

(a)
Time Data Forecasts Errors Absolute errors Squared errors

2 2
t X xA XB e
A e
B i e
A 1 1 e
B1 eA e
B
7 7 6 6 1.0 1.00 1.0 1.00 1.00 1.00

8 5 6.5 6.10 -1.5 -1.10 1.5 1.10 2.25 1.21
9 4 6 5.99 -2.0 -1.99 2.0 1.99 4.00 3.96
10 8 4.5 5.79 3.5 2.21 3.5 2.21 12.25 4.88
11 6 6 6.01 0.0 -0.01 0.0 0.01 0.00 0.00
12 7 7 6.01 0.0 0.99 0.0 0.99 0.00 0.98
13 7 6.5 6.11 0.5 0.89 0.5 0.89 0.25 0.79
14 5 7 6.20 -2.0 -1.20 2.0 1.20 4.00 1.44
Totals —0.5 0.79 10.5 9.39 23.75 14.26

Means —0.062 0.099 1.31 1.17 2.97 1.78
Root mean squares (RMS) 1.72 1.31
(b) Summary
Method A Method B
Mean error -0.06 0.10

Mean absolute error 1.31 1.17
(MAE)
Mean square error 2.97 1.17
(MSE)
Root mean square error 1.72 1.31
(RMSE)
mean error. Where this average error becomes large, perhaps due to
failure to follow accurately some systematic structure in the data, the
forecasting formula becomes suspect and needs adjusting or replacing
by another which is suitable for this particular type of structure. A
systematic deviation of the forecast errors from zero is referred to as a
‘bias’ in the forecasts. The mean on its own is not enough, since it may
well be close to zero while the actual errors are very large, positive and
negative errors tending to cancel each other out. To eliminate the sign
for the purpose of measuring spread, one may either just ignore it and
consider the absolute errors, denoted by I et I, or remove it by squaring.
The first approach measures spread by the mean absolute error (MAE),
1 "
MAE = - I | et\
n i= i
sometimes also called the mean deviation of error. The second approach
uses the mean square error (MSE),
31
MSE = — 2 ej
n i= 1
Table 3.4 gives the calculation of these quantities. It is seen that both
quantities indicate quite clearly that method B gives errors having a
smaller spread than method A.
To make the units of the mean square error the same as those of the
mean error and the mean absolute error, its square root must be taken
to give the root mean square error (RMSE). The choice of whether the
mean absolute error or the root mean square error is used is based on
practical considerations of the ease of calculation and the use to be
made of the calculated measures. It should also be noted that the root
mean square error places greater emphasis on the large errors than does
the mean absolute error.
If a bias is found to exist, it may be advisable to measure the spread
of the errors about their average rather than about the zero value. Thus
the mean square error would be replaced by the ‘sample variance’
defined by
2
1 n
s = — 2 (e — e)2
n i=i i
A little calculation will show that this is related to e and the mean
square error by
s2 = MSE -c2
The square root of this is called the sample standard deviation, s.
In many of the computer programmes available commercially and in
much of the literature on forecasting, the root mean square error is
used as the main criteria of forecast success. It is seen that
MSE = e2 +s2
Thus the mean square error is the sum of two contributory factors, one
measuring any bias there might be in the errors and the other the
variability of the errors about the mean error. An examination of the
mean square error without taking e into account may thus be
misleading or at least not as informative as a separate study of e and s2
or, equivalently, e and MSE.
In the same way that the mean square error can be modified to give
2
s , we may modify the mean absolute error by replacing the errors by
their deviations from the mean error, to obtain the mean absolute
deviation of errors (MADE),
1 n
MADE = - 2 \et — e\
n i= I
We are now in a position to say that from a statistical point of view the
32
simplest criteria of good forecasts are the requirements of small average

error and small mean square error or mean absolute error. These are
sometimes sufficient in terms of forecasting criteria, but in testing out a
forecasting formula on any reasonably extensive set of data it is
advisable to study the forecast errors in some more detail. This is of
value in giving a better ‘feel’ for the methods and data that are under
study, and also possibly leads to ways of improving forecasts.
3.4 Transient features

Transient features only become apparent where the data exhibit a
sudden change, a fracture, in their structure. The only way of
examining such features in data is to attempt to identify the occurrence
of such fractures and examine the behaviour of the forecast errors in
these regions. With clear fractures and well-behaved data, as in
Table 3.3 and Figure 3.2, the nature of the transients is clear. Table 3.3
exhibits the transient phenomena by simply removing the steady-state
values from the errors. With the usual messy data of the real world such
a procedure is difficult, if not impossible. Hence a study of the
transient features is best done using models of the types of structure
and fractures that either are shown by the actual data or might
reasonably be shown by that type of data. The examples and discussion
of section 3.2 show that in examining the success of a forecasting
formula the existence of large transient features in the forecast errors is
highly undesirable. Any effects that do occur should die away as
rapidly as possible. Thus from Table 3.3 this criteria is satisfied far
better by method A than by method B.
A type of transient feature of some importance in forecasting is due
not to a change in structure but to the way in which the forecasting
method is initiated. For example, some methods of forecasting require
an initial forecast to be made by the user before the forecasting process
is started on the actual data. If this initial forecast is a poor one, it will
influence the forecasts for some time to come.
3.5 Steady-state features

In analysing steady-state features in data, we meet very similar
problems to those in the previous section. One can examine the data to
find regions that appear to be far removed from any fractures in the
structure. This could avoid the problem of transient features. These
regions can then be studied to see if there is any clear steady-state
behaviour. The simplest way of doing this is to examine the average
error over these regions. As was seen in the example of Table 3.3, a
steady-state effect, such as lag, usually shows itself as a significant
deviation of the average error from zero. There are, however, other
33
types of steady-state behaviour for which the average error will be close
to zero. In particular, if an oscillation occurs in the data which is
followed by the forecasting formula with some lag, positive and
negative errors will occur in about equal proportions and thus a small
average error will occur; none the less, there is in this situation a clear
steady-state feature. The best and simplest method of examining such
features is to plot a chart of the errors against time. Chapter 15 gives
examples of such plots.
When dealing with a set of data we are very limited with what can be
done in investigating steady-state features, and hence again we must
turn to the use of models. Here we would use both models chosen to
describe the existing data and also models that describe features that
could well occur in the future structure. The analysis of the steady-state
effects, when the forecasting formula is applied to a model, is thus the
crucial test of the formula. The existence of steady-state errors is
obviously a disadvantage to any forecasting formula and it is necessary
to design formulae so that they are unlikely to occur. Where their
existance is acceptable, as a sacrifice made in order to obtain other
advantages, the effects should be kept as small as possible.
3.6 Criteria for testing forecasting formulae on models

In section 3.3 we based our criteria for forecast success on the actual
errors between our forecasts and our observations. From these errors
we calculated such things as the average error and the mean square
error. When we turn now to criteria as they apply to models, we have to
replace these numerical quantities by corresponding theoretical quan-
tities. To do this we make use of the statistical ideas of the expectation
and the variance of the distribution of errors. Table 3.5 summarizes the
definitions and formulae needed. As can be seen from the table, all the
quantities that we used in looking at the statistics of errors in data have
corresponding quantities for a population of errors arising from a
model. For the detailed theory of the expectation (sometimes called
the population mean) and variance the reader is referred to any
standard statistical text (e.g. see Hogg and Craig, 1970; Mood and
Graybill, 1963). We will limit ourselves to illustrating the ideas with
some forecasting examples.
Let us first apply the idea of expectation to a study of the transient
and steady-state errors that occur when a forecasting formula is used
with a particular model. For the sake of illustration we will use a model
which consists of a linear trend with an added random variation. Thus
we have for our observations xt,
xt = a + (3t + et
where the constant a. is the mean of the trend at time zero, /3 is the
34
Table 3.3 Expectation and related ideas
Distribution of errors P(e) = probability (error = e), for discrete

errors
or
f(e) = probability distribution function of
errors, for continuous case
Expectation of error (bias) E(e) = 2 eP(e), for the discrete case
all e
or
E(e) = f ef(e)de, for the continuous case
all e
Expectation rules E(ae + b) = aE(e) + b, for constants a and b

E(Xet) = 2£(c<)
i i
Mean square error for distribution MSE = E(e2)
Mean absolute error MAE = E(\ e |)
Variance of error Var(e) = E[{e —E(e)}2 ]
Variance rules Var(e) = MSE — E(e)2
= MSE — bias2
Var(ae + b) = a2 Var(e)
Var(2c/) = 2Var(cz), only if the errors et are
i i
independent
slope and the random variables et are independent observations on a

population with a normal distribution, mean zero, i.e. E(et) = 0, and
variance a2. Another way of describing this model would be to say that
the xt come from a normal population with mean a + fit and variance
o2. However, the above form, in which the deterministic and random
parts of the situation are clearly separated, is most useful for the type
of study we are undertaking. As a forecasting formula we will use the
forecast xt+1 ofxt+1 given by
A
/
V*
•**+1 J't
i.e. we simply use the last observation to forecast the next. This is the
simplest formula to use here for illustrative purposes. It corresponds to
the intuitive forecasts made by the type of individual who uses only his
own experience and feel for the situation as the basis for forecasting
but unfortunately has only a short memory. Having stated the model
and the forecasting formula, we need only to bring the two together to
obtain an expression for the forecast error. Thus,
(?t+1 ~ %t+ 1 %t+ 1
~ ^t+ 1
= (a: + P(t + 1) + ef+1} — (a + (3t + et}

35
Hence,
CJ + 1 =(3 + ef+, — e,
Using the properties of expectations given above, the expectation,
which is the average of the population of the errors, is
E (et+ j) =E(@ + et+1 — et)
= E (0)+E(et+1)-E(et)
=
P
since the E(et) = 0 for all t from the definition. Thus the forecast is
biased in the sense that the population average of the errors is f3, so that
on average the forecasts lag behind the observations by an amount (3.
This is a steady-state lag which remains the same over all time. To
illustrate a transient phenomena, consider what happens if at some time
T there is a sudden change in the situation which makes a become
ot + 5, this new value becoming permanent after time T. Since we have
shown that et does not depend on the value of a, the expression we
obtained for et will remain unaltered for times before and after eT. The
value of eT will, however, be different and we will have
& j' — Xj x j'
'V*
Jij1
/V*
T —1
= {ce + 5 + (3T + eT} — {cv /3 (T — 1) + 6T-_1}
=
j3 + 5 + 6 j 1
6 j1 _j
Thus the bias is as before, except at time T when the change in a
produces a transient effect in the form of sudden increase by 5 which
dies away immediately.
Let us now look at the spread of the errors. As we know the errors to
be biased, it is more useful to look at the variance than the mean square
error. This variance is the population quantity, from the model, which
corresponds to the sample variance s2, discussed in section 3.3. From
the above results we now have
Var(ef+ j) = Var(|3 + e,+ j - et)
= Var(ei+1) + Var(et)
= 2 a2
using the fact that et and et+i are independent with variance a2, and
letting b = P and a = — 1 in the formula for Var(ae? + 6). We see then
that the variance of the forecast errors is twice the variance of the
individual observations. This is a large error variance and suggests that our
forecasting formula would not be of much practical use for data from
this model. The mean square forecasting error is
E(e2) = 2o2 + P2
36
from the relation in Table 3.5 between the mean square error, variance
and bias.
We have now seen that by substitution of the model in the
forecasting formula we may investigate the main steady-state, transient
and statistical properties of a forecasting situation.
3.7 Study by simulation
The forecaster can investigate quite complicated formulae and models

using the ideas illustrated in the previous section. However, it
frequently occurs that formulae and models get too complicated for
this type of investigation or alternatively some information may be
required without going to all this mathematical trouble. One method of
study that can universally be applied is a study by simulation. To carry
out a simulation study, artificial sets of data are constructed using the
model, and the forecasting formula is then applied to these data. The
results of the forecasts are then analysed by the usual methods of
studying data discussed in sections 3.3, 3.4 and 3.5. For fairly simple
cases this can be done by hand quite easily. If random models are to be
studied in detail or a large number of possibilities are to be looked at,
then computer simulation becomes advisable.
The data in Tables 3.2 and 3.3 are simple examples of data forged to
show certain properties. Table 3.6 shows a working layout for
Table 3.6 Working layout for simulating a linear trend
xt = a + (3t + et a = 50
(3 = 2
o = 2, the standard deviation required
ritP\ N (0, 1) obtained from tables
of random normal deviates
t a fit Vt e* = xt = a + (3t
0 50 0 -1.085 -2.170 47.830

1 50 2 -0.277 -0.554 51.446
2 50 4 -0.251 -0.502 53.498
3 50 6 0.817 1.634 57.634
4 50 8 -0.123 -0.246 57.754
5 50 10 -0.002 -0.004 59.996
6 50 12 -1.007 -2.014 59.986
7 50 14 -2.226 -4.452 59.548
8 50 16 -1.580 -3.160 62.840
9 50 18 1.468 2.936 70.936
10 50 20 0.169 0.338 70.338
11 50 22 1.530 3.060 75.060
12 50 24 0.598 1.196 75.196
37
simulating data from a trend model. The table shows one run of the
simulation using one set of es. Sets of random normal deviates, as such
quantities from a normal distribution iV(0,l) are called, may be
obtained from tables (e.g. Rand Corporation, 1955) or generated by
standard instructions on most computers. The simulation of data with a
seasonal variation was illustrated in Table 1.3.
References
Hogg, R. V. and Craig, A. T. (1970). Introduction to Mathematical Statistics.
Macmillan, New York.
Mood, A. M., and Graybill, F. A. (1963). Introduction to the Theory of Statistics.
McGraw-Hill, New York.
Rand Corporation (1955). A Million Random Digits with 100,000 Normal Deviates.
The Free Press, Glencoe, Illinois.
'
Part II
Forecasting for some basic models
In the next eight chapters we will examine a range of models that form
the basis of most practical forecasting. The models discussed are
The constant mean model Chapter 4
The linear trend model Chapter 5
Regression models Chapter 6
Stochastic models Chapter 7
Seasonal models Chapter 8
Growth curves Chapter 9
Probabilistic models Chapter 10
Multivariate models Chapter 11
For each class of model we will examine various ways of obtaining
forecasts. In so doing we will develop a range of methods, principles
and approaches that have wide applicability in all types of forecasting;
these are discussed more generally in Chapter 12.
Chapter 4
The constant mean model
4.1 The global constant mean model

The form of this model is
xt = p + et (t = 1, 2, . . .)
where p is the constant mean and et is one of a sequence of
independent random variables with zero expectation and constant
variance o2. Table 1.1(d) showed some data from such a model which is
characterized by a random variation about the constant level p. We
assume that observations xly x2, . . . , xt have been obtained and a
forecast of a future observation xt+k is required. We shall denote the
forecast by xt> k; the first subscript, £, denotes the time at which the
forecast is obtained and the second subscript, fe, denotes the time into
the future which we span with our forecast, the lead time. The actual
future value xt+k can be expressed in terms of our model by
%t+k ~ T + €t+k
Thus to forecast xt+k we shall have to assign values to both p and et+k.
The random variable et+k is by definition independent of the available
information contained in xl9 ... 9 xt and hence we cannot give a
forecast of its future value, save to note that its expected value wall be
zero. So we simply forecast et+k by the value zero. The quantity pis,
for our global model, a constant over all time, and therefore an estimate
of p at the present time will also estimate its future value. The most
natural estimate of the mean p is the sample mean (average) based on all
available data. Combining this with the zero forecast value for et+k
gives as our forecast the expression
Xfk (*^ l *^2
We shall term this forecasting formula the ‘explicit’ form of our

forecast, since it explicitly includes all our data.
The right-hand side of the above expression does not depend on k
41
42
because a constant mean is assumed and our observations are in-

dependent of each other. For the rest of this section we will drop the
subscript k. It can be shown that, under certain conditions, the above
forecast minimizes the mean square forecasting error, that is to say
E(xt+k— xt)2
is minimum for this particular forecasting formula. As observations are
usually obtained on a continuing basis, it is useful to try to put the
forecasting formula in what is called recurrence form, i.e. in a form in
which we can obtain the new forecast from the old forecast and the
newly obtained data with the minimum of calculation. The new
forecast %t+ , which will use the new observation xt+ \, can be related
to the previous forecast in a quite simple fashion. Thus
%t+ 1 = [(*1 + *2 + • • • + Xt ) + Xt+ i ] /(£ + 1)
and hence
Xt+1 = [txt + Xt+l ]/(f + 1)
Thus instead of having to keep a record of all past data, we need keep
only the value xt. When our next observation xt+l is obtained, the
above equation provides a simple way combining it with the last
forecast xt to obtain the next forecast xt+l. Table 4.1 gives a simple
illustration of such a calculation, re-expressing the formula to give xt in
terms of xt_ ! and xt.
If we regard xt as the forecast of the next observation, x 1, then we t+
can express the forecasting formula in yet another form. The basis of
this third form is the use of the forecast error. Denote the forecast error
at time t by et, where
Xf xt — I
Table 4.1 Forecasts for the global constant

mean model — recurrence form
t—1 1
Xt x
f-I
+ xt, x0 = 0
i
t—1
t xt xt/t xt-1 xt-1 xt
t
1 7 7.000 0 0 7.000
2 3 1.500 7.000 3.500 5.000
3 9 3.000 5.000 3.333 6.333
4 6 1.500 6.333 4.750 6.250
5 8 1.600 6.250 5.000 6.600
6 7 1.167 6.600 5.500 6.667
43
Table 4.2 Forecasts for the global constant

mean model — error correction form
5ct = £f-i + e,/t, et=xt- x
t- I > xo = 0
t xt Xt- 1 e
t et/t xt
1 7 0 7.000 7.000 7.000

2 3 7.000 -4.000 -2.000 5.000
3 9 5.000 4.000 1.333 6.333
4 6 6.333 -0.333 -0.083 6.250
5 8 6.250 1.750 0.350 6.600
6 7 6.600 0.400 0.067 6.667
Then substituting for xt+1 in the last formula gives

xt+i = (txt+e ,+1 + xt)/(t + 1)
Hence,
X t+ 1 ~ ^ t + &t+ l 1)
Our new forecast is obtained from the old simply by adding a

multiple of the last forecast error. A simple example is given in Table
4.2. The multiplying factor is 1/number of observations, which gets
smaller and smaller as t increases. Thus the forecast gets less and less
sensitive to forecast errors. A large change in structure could produce
large errors, but yet because of the size of the divisor the forecast
would be very little changed. The last formula is called the error
correction form.
In the above discussion we have expressed our forecasting formula in
three forms:
(a) the explicit form;
(b) the recurrence form;
(c) the error correction form.
We shall see in following chapters that these three forms not only
represent different ways in which we can express the same forecasting
formula but can also provide us with completely distinct approaches to
the development of forecasting methods.
Having discussed different forms of the forecasting formula, we must
next examine its properties. To do this, we simply make use of the
model and of the simple rules that were given in Chapter 3.
Using the model we can re-express the forecast as
Xt = 2 Xj/t
t= 1
= 2 (;u + e,)/f
t= 1
44
Hence,
Xt = ju + 2 e,/f
?= 1
Taking expectations of both sides and using the facts thatE(ju) = p and
E(€i) = 0, we have
E(xt) = n
Thus the forecast is unbiased. The forecast error is
^t,k ~ Xf+ k Xf k
using an obvious notation, so

t
+
et,k = H et+k —fi — 2 e,-/t
i- 1
t
~ ^t+ k ^ ft
i= 1
As E(et k) = 0, for all k, the forecast is thus again proved unbiased for
all future values.
Finding the variance of both sides of the above expression for x t and
using the facts that Var(p) = 0 and Var(ei ) = a2, for all i, together with
the independence of the e values, gives
Var(x;) = Var^2ê,-/f j
= 2 Var(edit2
i= 1
Hence,
Var(xr) = o2ft
Since the mean error is zero, as the forecast is unbiased, it follows that
the mean square error is the same as the variance of error. Hence,
MSE = Vax(xt+k -xt)
But xt+k and 3ct are independent of each other and so
MSE = Var(xf+/C) + Var(xf)
= o2 + o2 ft
t
45
Thus as t gets large the forecast estimates p with high precision and the
mean square error approaches o2, which is due simply to our inability
to forecast the random variable et+k.
Knowledge of the fact that the forecast is unbiased with error
variance (mean square error,)
enables us to make much more useful forecast than the bare statement
that the forecast is such and such a number. In particular, we can make
statements about the probable range of values to be taken by the future
observation. The most common form for such statements is called a
confidence interval. We will outline the basic idea, but for details the
reader is referred to any standard statistical text, e.g. Hogg and Craig
(1970) and Mood and Graybill (1963).
Let us start by assuming that the forecast errors are normally
distributed with zero mean and variance o^. It follows from the
properties of the normal distribution that the probability of the
forecast error lying in the interval -—1.96 oe and +1.96 oe is 0.95. In
general, the probabilities associated with ranges of values of a random
variable are given by the areas under their distribution curves. In the
normal error distribution (see Figure 4.1) these are determined by the
spread of the curve as measured by oe. If we require different
probabilities, then the 1.96 is replaced by some other appropriate
Interval containing
observed x on 95%
of forecasts
Figure 4.1 Prediction intervals

number obtained from tables of the normal distribution (e.g. see Fisher
and Yates, 1938). We may write our statement as
Prob (— 1.96 oe < e < 1.96 oe) = 0.95
Since
t + 1
et+i = xt+ i — x t and o\ = cr
this may be simply rewritten as
\
t+ 1 t + 1
Prob —1.96 .o<xf+1 —£*<1.96 a \ = 0.95
and rewritten again as
Problx* —1.96 j—~—o <xt+ ! <oct + 1.96|=0.95
Notice that, assuming a is known, the terms forming the limits of the
inequality are all known and can be calculated as two numbers. These
we call the lower confidence limit (LCL) and the upper confidence
limit (ULC). Thus,
Prob (LCL < x*+ x < UCL) = 0.95
The nature of this change from e to xt+i is indicated in Figure 4.1.

Care must be taken over the interpretation of such probability
statements. The probability does not refer to the value of xt+ 1, for this
particular set of data, but rather it indicates that, in the long long run,
using this method repeatedly will yield values of LCL and UCL which
on 95 per cent of occasions will lie to either side of the observation
being forecast. By way of example, suppose we have eight weeks’ data
believed to come from our model with o = 10 and the forecast xt is 245
items. The two confidence limits are
/9
LCL = 245 - 1.96 / - x 10 = 224.2
V 3
/9
UCL - 245 + 1.96 / - x 10 = 265.8
V 8
Thus the 95 per cent confidence limits for the forecast are 224.2 and
265.8. The interval 224.2 to 265.8 is called the 95 per cent, confidence
interval, or prediction interval.
Notice that, though for ease or notation we have talked about
forecasting xt+!, we could have talked of forecasting xt+k without any
change in the calculations. Similarly, we have treated the observations
x{, . . . , xt as though they were made at equally spaced times called 1,
2, . . . , t. The observations could have been made at any times, say Tx,
47
T2, . . • , Tt, without altering our results. The reason for this is that in
this model the basis of the forecast is the estimate of \u which is
assumed constant for all time. Thus it does not matter when the
forecasts are made for or when the observations were taken. In all cases
we get the same results and, in particular, the same confidence interval.
We will see in later sections that this is not true in general.
In the previous discussion it has been assumed that a is known,
which is obviously very rarely true. When any parameter is not known,
it must be estimated from the observations. To estimate o2, the
variance of x, we use a simple modification of the sample variance of
the data, which was defined in section 3.2. Here we define our estimate
as
1 t
2 (x,—x)2
where x is the mean of the available data, which is xt for the model of
this section. If this estimate is used in place of a2 in the calculation of
confidence limits, the value 1.96, or whatever was used from the tables,
has to be replaced by the corresponding value from the table of the
^-distribution. The reason is that when the constant a2 is replaced by
the random variable a2, which depends on the data, the normal
distribution form of Figure 4.1 is no longer applicable and a
distribution called a ^-distribution is used instead (e.g. Fisher and Yates,
1938). If in our example the value of a2 was 100 and the value of
t — 1, called the degrees of freedom, was 7, then looking up the 95 per
cent value for 7 degrees of freedom in tables of the f-distribution gives
the number 2.365 in place of 1.96. Recalculating the confidence limits
gives
LCL - 245 - 2.365 /- x 10 = 219.9

V 8
/9
UCL = 245 + 2.365 / — x 10 = 270.1
V 8
Thus the prediction interval is wider, which simply reflects the fact
that, when a is not known, we have less knowledge of the situation and
cannot give as short an interval as before. In the above calculation we
have obtained the error standard deviation oe from that of the data a.
As we will usually measure the errors directly it is more natural, and
reliable, to estimate oe directly from the observed errors. We would
thus use
1 t
-2c
t i=
Table 4.3 Global constant mean model
Model xt = ii + et, t = 0,1,. . . , et independent, E(et) = 0,

Var( et) = o2
Data • • • » ^t
Minimum MSE
forecast of xt+k is xtk = xt independent of lead time k
-“ M *'*■
xi
Explicit form
X
II
h
Recurrence form xt = Xt_i(t — 1 )lt + xt/t
Error correction form xt = xt_1 + et/t, et = xt-xt_l
Forecast bias 0
Forecast MSE (t+\)o2/t
95% prediction l t + 1
1.96 a // ,xt+ 1.96 a /A +1 \)
interval, o2 known
The prediction interval would then be between x ~ ctbe and 5c + ctde,

where ct is the appropriate constant as given in tables of the
^-distribution, using t as the ‘degrees of freedom’.
The importance of attempting to give a confidence interval for a
forecast cannot be exaggerated. To state, in the above example, that the
forecast is 245 items, without any indication of its possible accuracy, is
of little use. In the example, the 95 per cent confidence limits were
(224.2, 265.8) when o was known. Suppose, however, they had been
(0, 490); then our attitude to the forecast of 245 would have been
entirely different. Table 4.3 gives a summary of the main results of this
section. Notice that in xtk all observations carry equal weight. The
earliest observation, xl9 is as important as the latest, xt. This is a
consequence of the fact that we have treated our basic model as being
globally valid. We have used it as a global constant mean model.
4.2 The local constant mean model

The basic structure for the local constant mean model is exactly that
given in the last section. However, it is now believed that, though this
model is a good one for short periods of time, it is unlikely that the
value of ju for one locality in time is exactly the same as that for
another. Thus p is regarded as a slowly varying quantity. Later we shall
49
Table 4.4 Calculation of moving averages
Month 1 2 3 4 5 6 7 8 9 10 11 12
Sales 7 5 6 4 5 4 5 3 4 7 6 5
First set 7 5 6 4 5
Average 5.4
Second set 5 6 4 5 4
Average 4.8
etc.
Eighth set 3 4 7 6 5
Average 5.0
Moving averages 5.4 4.8 4.8 4.1 4.2 4.6 5.0 5.0
consider ways of modelling this variation in p, but for now we shall

assume that it might happen in a slow but unspecified way. In the
global constant mean model the averaging operation over all the data
had the effect of cutting down the random variation and leaving an
estimate of p. If an estimate of p is required in just one locality of the
data, then it is reasonable to average the observations just from that
locality and ignore the rest. In Table 4.4 some illustrative figures
indicate how an average can be obtained from any five-month period in
the total set of data. Thus if we wish to estimate p for the most recent
period of, say, 5 months, we simply average the last five observations.
Hence the estimate is (3 + 4 + 7 + 6 + 5)/5 = 5.0. This would, at this
simple level, be the estimate of p that could be used to forecast xx 4. If
the mean, p, was wandering to any extent in the time covered by the
five observations, this forecast would obviously be a biased forecast. In
this situation all we could say would be that the value of the average,
5.0, is an estimate of p at the middle of the five observations, i.e. at
week 10. Similarly, the averages of each successive group of five can be
regarded as estimates of the mean at the centre of their ranges. Thus the
last line in Table 4.4 gives the estimates of the means for each month
found by moving the sequence of five numbers successively to the right
by one. The averages so obtained are thus termed ‘moving averages’. It
will be seen that there are less moving averages than observations, and
the latest value of the mean estimated will always be that corresponding
to some time before the latest observation. Where an average of 2n + 1
observations is used, there will be n observations after the time at which
the latest moving average is centred. If there are an even number of
observations, the centre of the range of the moving average will fall
50
between two observations. To obtain values that estimate the means at

observations times, these moving averages are themselves averaged in
pairs.
If a moving average of s observations is used, the forecast of xt+k
would be
x + +
t-s+ 1 • • • Xt
Xf =
s
When new observations are obtained, a recurrence relation can be used
since
x
t+ 1 ~ (*^?+ 1 &xt x
t — s+ 1 )/^
Both these forms involve storage of the last s observations. If a moving

average is to be seriously considered as a means of forecasting, there
must be good evidence that the mean is only wandering slightly. A
somewhat different use for moving averages in forecasting will be
considered in a later chapter.
As already pointed out, a moving average is essentially an estimate of
the mean at the mid-time of the time interval covered. Its main
practical use is thus to smooth the data rather than forecast the future.
As pointed out by Davis (1941), when the interest is in forecasting it
would be more reasonable to use a moving average which gives more
weight to recent data than past data. For example, instead of a simple
moving average
1
— (*,_3 + Xt_2 +Xt_ ! +*,)
4
we might use
+ 2xt_ 2 + 3xr_ i + Axt)

10
This would give four times as much attention to xt as to xt_ 3, instead
of the equal attention given by the ordinary moving average. This
would intuitively make it more suitable for forecasting purposes. The
divisor, 10, is introduced as being the sum of the weights, 1 + 2 + 3 + 4.
The reason for this is that if, for example,
X f— 3 1 — 2 f— 1 Xf I
then this ‘weighted’ moving average will also have the value 7. Though
this weighted moving average estimates p locally, in the sense that it
takes most notice of current information, it suffers from two
disadvantages. Firstly, it only uses the four latest observations and
ignores the rest. Secondly, it does not possess a very simple recurrence
form. A set of weights for a weighted moving average which overcomes
51
both these disadvantages is the ‘exponential’ set of weights given by

Observation OC j X2 X^ 2 »X ^j 0C> ^
1 0
Weight at~2 a2 a1 a
where a is a constant with value 0 < a < 1. The general formula for the
exponentially weighted moving average is
t- I
Z arxt_ r
r= 0
Xt =
t- 1
Z ar
r- 0
Table 4.5(a) shows the calculation of this weighted mean directly as in

the formula. This looks very complicated until one realizes that each
week’s new values can be calculated from the previous week’s values.
Thus the weighted sum of the observations, the numerator, for week 3
is
0.82 x 7 + 0.8 x 5 + 1.0 x 6
which is
0.8 (0.8 x 7 + 5) + 6
which is
0.8 (weighted sum for week 2) + latest observation
similarly, the sum of weights, the denominator, for week 3 is
0.82 + 0.8 + 1.0
which is
0.8 (0.8 + 1.0) + 1.0
which is
0.8 (sum of weights for week 2) + 1.0
We can thus replace the above formula by a simple set of calculations as
follows. Denote the numerator of xt by St and the denominator by Wt,
i.e.
t- I
St = Z arxt_ r
r= 0
and
1
Wt = f Z ar
r= 0
52
Table 4.5(a) Calculations for exponentially weighted mean,

explicit form a = 0.8
Week number 1 ' 2 3 4
Week Sales 7 5 6 4
1 Weight 1.0
Weighted sum 1.0 x 7 = 7
Sum of weights 1.0
JCi 7/1.0 = 7
2 Weights 0.8 1.0

Weighted sum 0.8 x 7 + 1.0 x 5 = 10.6
Sum of weights 0.8 + 1.0 = 1.8
x2 10.6/1.8 = 5.9
3 Weights 0.82 0.8 1.0

2
Weighted sum 0.8 x 7 + 0.8 x 5 + 1.0 x 6 = 14.48
Sum of weights 0.82 + 0.8 + 1.0 = 2.44
x3 = 14.48/2.44 = 5.93
Table 4.5(b) Tabular layout for exponentially weighted mean, a = 0.8
a) (2) (3) (4) (5) (6)
Time Observation 0.8 x previous (2) + (3) Divisor (4)/(5)

t s,
M
xt s xt
II
1 7 0.0 7.0 1.0 7.0

2 5 5.6 10.6 1.8 5.89
3 6 8.48 14.48 2.44 5.93
4 4 11.58 15.58 2.95 5.28
5 5 12.47 17.47 3.36 5.20
6 4 13.97 17.97 3.69 4.87
7 5 14.38 19.38 3.95 4.91
8 3 15.50 18.50 4.16 4.45
9 4 14.80 18.80 4.33 4.34
Then
St - aSt_ x + xt
Wt=aWt_ i + 1
xt = St/Wt
s0=o,w0=o
Thus both numerator and denominator may be evaluated using very
simple recurrence formulae. This is illustrated in Table 4.5(b). In
53
moving from one stage to the next of this calculation, only the last
values of St and Wt need to be recorded.
One way of looking at the introduction of the weights ar is that we
now look at our data through a fog. When a is small the fog is thick, so
that we can only see with any clarity the data that are close to us in
time. Putting a = 1 corresponds to a clear day in which we can see all
the data clearly. Technically, the association of weights with variables
to decrease their importance is called discounting and we will refer to a
as the discounting factor.
Let us now consider two other approximate ways of expressing our
forecasting formula. If a < 1 and if t is so large that af is negligible, then
the value of Wt can be shown to tend towards a constant limit of
(1 — a)-1 and thus only St needs to be stored. In this case the formula
becomes
t- I
xt - (1 — a) 2 arxt_ r
r= o
This can be rewritten in two ways that are very convenient for
calculation purposes:
xt = (1 — a)(xt + axt_ ! + a2xt_ 2 + . . .)

= (1 — a)xt + a(l — a)(xt__ x + axt_ 2 + . . .)
Thus
xt = (1 — a)xt + axt_ i.
This recurrence relation, as that for the ordinary average in section 4.1,
relates the new forecast to the old forecast and the new observation.
The values a and 1 — a give the proportion of weight attached to the
old forecast and new observation, respectively. In words, we may thus
write
New forecast = (1 — a) x new observation + a x old forecast

This recurrence relation can be treated as a forecasting formula in its
own right, instead of as an approximation, and is referred as
‘exponential smoothing’ (e.g. Brown, 1959; Winters, 1960). The fact
that the weights are constant enable a simple nomogram to be used for
updating. This is shown in Figure 4.2.
We may also rewrite the recurrence equation in terms of forecast
errors. Writing
e™ 'V* y
t ^t t— l
the last equation becomes
xt = (1 — a)(et + xt_ i) + axt_ x

54
Figure 4.2 A nomogram for exponential

smoothing
Hence
x, =xt_1 + (1 — a)et
Thus the new forecast can be obtained from the old by adding a
fraction (1 — a) of the last error made in our forecasting. The only
difference between this and the similar expression for the global mean
model is that the fraction is a constant independent of t rather than a
fraction which gets smaller as t increases. Intuitively, this means that
the forecast is always ‘constantly’ alert for changes in the situation
which would reveal themselves through the forecast errors. This is
exactly the feature that is required if we are to forecast a constant
mean model which is valid locally rather than globally. The method has
been introduced to deal with a model approximating to the global
constant mean model
xt = n + et
over small localities of time, but for which the mean ju wanders in some
unspecified fashion. Many theoretical models have been devised that
give, in effect, wandering means. It has been shown that for many of
these the exponentially weighted moving average gives a good forecast.
Clearly, if we can be more precise about how the mean wanders, we
should be able to improve on this forecast. We will discuss suitable
methods in later chapters. Experience shows, however, that this
method, often called exponential smoothing, provides a very robust and
widely applicable technique.
55
As a means of studying the method, let us compare it with the

method discussed in the previous section, assuming the ‘true’ model to
be the global constant mean model. Consider the expectation and
variance of xt- St/Wt■ For a geometric series, which our so-called
‘exponential’ weights actually form,
Wt = 1 + a + a2 + . . . + af~ 1
= (1 — at)/( 1 — a)
and from the model

t- I t- I
r
St = p 2 a + 2 aret_ r
r =0 r=0
Hence
E(St) = nWt + 0
and
E(xt) = n
Thus exponential smoothing leads to an unbiased forecast if the mean is
genuinely constant. Further,
2t
t- I 1 — a
Var{St) = 2 a2ro2=~ y o2
r— 0 1 Cl
But
Var(x?) = Vax(St)/Wt2
and so after simplifying
1 —a 1 + af
Var(x,) * . a
1 +a 1 — af
For t large this becomes
Var(xf) = a2 (0 < a < 1)

1 +a
This variance is fundamentally different in form to that obtained for the
forecast in the previous section, o2 /£, which becomes smaller as t is
increased and in fact approaches zero. For exponential smoothing the
variance of xt approaches the minimum value given above and however
much data we have we shall never get a smaller variance than this. The
reason is intuitively clear since as new data are added less and less notice is
taken of old data, so the effective amount of data available remains
constant when t is large. It will also be noticed that as a, the
discounting factor, gets closer to one, for any fixed t, Var(x) gets
smaller and a plot of xt will look smoother. This occurs because, as we
56
Table 4.6 Some values of Var{xt)jo2 for constant mean models
t = 10 t = oo
Global Local Global Local

model model model model
0.10 1.27 0 oo
xt - average of x s
xt = exponential smoothing 0.14 0.19 0.11 0.12
saw in the previous section, the best forecasting formula for this model
is the ordinary average which corresponds to the exact form of
exponential smoothing with a = 1. By comparing values of Var(x) for
various values of a, including a- 1, we can see howr much we lose by
using our exponential weights when we are in fact dealing with a global
constant mean model. Conversely, if we consider some model for a
wandering mean process and repeat the comparison, we can see the
advantage of introducing the weights ar. Table 4.6 shows such a
comparison for a particular wandering mean process. Notice from the
table how little is lost, by way of having a larger variance, when a = 0.9,
say, is used instead of a = 1.0 in the global constant mean model,
and conversely how much bigger the variance is if a = 1.0 is used when
some a < 1.0 should have been used. It would appear from this example
that unless we can be absolutely certain that the constant mean model
has applied in the past, and will continue to apply in the future, then it
will be safer to use the weighted mean rather than the ordinary mean.
In the previous section we discussed the construction of confidence
limits for the future observation. It might appear reasonable to use the
above expressions for variance to construct such limits, and indeed we
could. The limits, however, would refer to the permanent model; the
limits for a local model would, in general, be different. The natural
approach here is to estimate oe directly from the observed forecast
errors using the estimator be given in the last section. The prediction
Table 4.7 Local constant mean model — exponential smoothing
Model xt = jU + et, t = 0,1,. . ., et independent, E(et) = 0, Var(ef) = o2

Model regarded as a local approximation
t-1 n-\
Explicit form xt = X a xt_r/' 2 ar, independent of lead time k
r
r-0 / r= 0
Recurrence xt = (1 — a)xt + axt_l, t large, a < 1

form
Error correction xt = xt_1 + (1 — a)et, t large, a < 1
form
57
intervals can then be derived using the same approach as that discussed
in that section.
The forecasting formula of exponential smoothing involves the
‘discounting factor’ a, which is a ‘forecasting parameter’ as distinct
from a parameter of the model. The value of a is normally chosen so as
to minimize the mean square error of forecasts over a trial period. This
is discussed in more detail in Chapter 16. Table 4.7 gives a summary of
the main formulae for this section. It should be noted that much of the
literature on exponential smoothing uses a = 1 — a as the basic
parameter. This is often referred to as the ‘smoothing constant’.
References
Brown, R. G. (1959). Statistical Forecasting for Inventory Control. McGraw-Hill,
New York.
Davis, H. T. (1941). The Analysis of Time Series. Cowles Commission, Yale.
Fisher, R. A. and Yates, F. (1938). Statistical Tables. Oliver and Boyd, Edinburgh.
Hogg, R. V. and Craig, A. T. (1970). Introduction to Mathematical Statistics.
Macmillan, New York.
Mood, A. M. and Graybill, F. A. (1963). Introduction to the Theory of Statistics.
Winters, P. R. (1960). Forecasting sales by exponentially weighted moving averages.
Man. Sci., 6, 324—342.
Chapter 5
Linear trend models
5.1 The global linear trend model

We turn now to the problem of forecasting future values of a structure
which shows a linear trend with added random variation. Thus the
model is
xt = a. + pt + et
where a: is the expectation of x0, (3 is the slope and e is a sequence of
independent random variables with F(ez) = 0, Var(e) = o2 for all i, and
whose distribution will, for the moment, remain unspecified. The data
in Table 5.1, charted in Figure 5.1, are typical of data for which such a
model might be reasonable. If this data did arise from the model, then
the model describes the future value x { 2 as
x12 = a + ]3.12 + el 2
As we have assumed the es to be independent of each other, we

have no way of forecasting e12. All we can do is to set e12 at its
expected value of zero. Thus if we knew a and (3, our forecast would be
a + j3.12. Unfortunately, we do not know a or j3 so we must use our data
to find values for them. In this section we shall investigate several ways
of estimating the values of a and (3 from a set of data and hence of
extrapolating the line to obtain forecasts of a future xt.
5.1.1 Fitting by eye

When a set of data is replaced by a corresponding curve, we say that
we have fitted the curve to the data and the curve is called the fitted
curve. In fitting a straight line to data such as that of Figure 5.1 we can
simply lay a transparent ruler over the plotted data and adjust it until
the data is scattered evenly and approximately equally about the edge
of the ruler. Figure 5.2 shows such a line for the data of Table 5.1. The
ease with which we can do this depends on how close the data looks to
58
59
Table 5.1 Data from a global linear trend model
Time 1 2 3 4 5 6 7 8 9 10 11
31 34 37 36 42 39 42 45 47 47 52
Figure 5.2 Fitting a line and extrapolating

60
a straight line. This is obviously a very easy way of obtaining a line. It

has the advantage that if we know that certain of the points are suspect,
in the sense of not being known as accurately as the others, or that they
are produced by known and exeptional events, then we can take less
notice of these points in assessing the line. However, this subjective
element is also the weakness of the method for, in data where the
variability of the random component is large, different people may well
produce very different lines. Further, as we have emphasized before, it
is always necessary in forecasting to know as much as possible about
the forecast obtained from a method — in particular, it is essential to be
able to specify, under the assumption that the model continues to be
valid, an estimate of the mean square error of the forecasts. The
subjective nature of fitting a line by eye means that there is no built-in
means of assessing such statistical properties. We can obviously examine
the statistical properties of the distances of the observations from the
fitted line to see if they are reasonable. For example, in Figure 5.2 we
could measure the distances, e:, . . . , et, called the residuals, of the
points from the fitted line, and see whether the average of these was
close to zero and look at the size of their mean square. However, by the
time we have done all the calculations to obtain such results we might
just as well have fitted the line by some more rigorous method. Such a
method is the method of least squares.
5.1.2 The method of least squares
Refer again to Figure 5.2 which shows a fitted line and the data. To
decide on how to fit a line, we must first decide which criteria the
fitted line should satisfy. The intuitive criterion is simply that the line
should be as close as possible to the observations. That is to say, that
the residuals ex, . . . , et should be small. By analogy with the criteria
for forecasting developed in previous chapters, an obvious criteria
would be that the mean square error of the ef should be as small as
possible. Since t is here fixed, this requirement is that the quantity S,
where
s= 2 if
i~ 1
is as small as possible. This criterion is referred to as the least squares
criterion. From the definition we can see that if xt is the observation
made at time i then
€j = xt - a — pi
where a and $ are fitted values, the estimators, of a and (3 respectively.

61
Thus
S = 2 (xt — a — $i)2
i= 1
A
The least squares line is the line for which a and (3 makes 5 a minimum.
A direct application of calculus (setting dS/da= 0 and dS/d(3 = 0) gives
two simultaneous equations for a and j3. These are
t A t
2 xt = at + (3 2 z
i= 1 i= 1
2 iXj = a 2 i + (3 2 z2
f=l i= 1 i=l
These equations are referred to as the ‘normal’ equations. If the times

at which the observations are taken are not simply 1, 2, . . . , t, then the
summations refer to all the times at which the actual observations were
made. Notice the relation of these equations to the equation of the
fitted line:
(a) The fitted line is x* = a + jSi;
(b) summing both sides over all observations gives the first normal
equation;
(c) multiplying both sides by the time, z, and then summing gives the
second normal equation.
It sometimes helps in solving these equations to use artificial time
origins. Table 5.2 gives an example in which i = 6 is taken as a new time
origin from which time i is measured. As 21 = 0, the first normal
equation becomes a = x, the average value of x, and the second gives
{$ - 'Ll x/'Ll 2. Note that here a: is the expected value of x at time i = 0.
Croxton, Cowden and Klein (1968) discuss various ways of simplifying
the solution of normal equations. Notice that if we know (3 to be zero,
for the constant mean model, the result a = x shows that the estimate
of p used in the previous chapter was in fact the least squares estimate.
If we keep to an origin for which the sum of the observation times is
zero, i.e. i = i — average time z, and we substitute for xt the value
a + (3i + e/, from the model, then we can show that:
(a) E(&) = a,E(P) = (3;
(b) Var(o:) = o2 /f, Var(|3) = a2/2z'2, t still being used to denote the
number of observations;
A
(c) a and (3 are independent.

The forecast of any future value of x, say xT, is simply obtained by
substituting the time T in the fitted line; thus
xT = a + (3T
62
Table 5.2 Calculation of least squares line

.r • >2 i
i X l l l X a + pt’ e e2
1 31 -5 25 -155 31.68 -0.68 0.462

2 34 -4 16 -134 33.56 0.44 0.194
3 37 -3 9 -111 35.44 1.56 2.434
4 36 -2 4 -72 37.33 -1.33 1.769
5 42 -1 1 -42 39.21 2.79 7.784
6 39 0 0 -514 41.09 -2.09 4.368
7 42 1 1 42 42.97 -0.97 0.941
8 45 2 4 90 44.85 0.15 0.023
9 47 3 9 141 46.74 0.26 0.068
10 47 4 16 188 48.62 -1.62 2.624
11 52 5 25 260 50.50 1.50 2.250
721
452 0 110 207 22.916

!
Xj' = a + Pi a = x = 41.09
Si' = 0 i$ = Xxi'/£i'2 = 1.882
•
Fitted line := 41 .09 + 1.882 (i--6) = 29. 80 + 1.88 i
Estimated variance b2 = 2e2/9 = 1.592
Forecast of x\2 = 41.09 + 1.882 x 6 = 52.38
The forecast error is
aT — Xj1 Xj'
= {a + PT + eT)~ (a + pT)
= a - a + (p - p)T + eT
Hence E(eT) = 0, from (a) above, so the forecast is unbiased. It follows

that the mean square error,
M.S.E. = E{e2T)
= Var(er)
= Var(&) + T2 Var(|3) + a2
Thus
/ 1 T2 \
M.S.E. = o (j + 2
+ l)
It will be seen that as T, the time from the average time of the
observations to the time being forecast, increases so does the mean
63
Table 5.3 Forecasts and limits for data of Table 5.1
(a) Known o2 = 2.5
Time Xf 95% prediction interval
12 52.380 1.883 48.689 to 56.071

13 54.266 1.960 50.424 to 58.108
14 56.148 2.045 52.140 to 60.156
15 58.030 2.137 53.841 to 62.218
16 59.912 2.237 55.527 to 64.297
(b) Estimated d2 = 2.546 using 9 degrees of freedom
Time Xy 95% prediction interval
12 52.380 1.900 48.082 to 56.678

13 54.266 1.978 49.792 to 58.740
14 56.148 2.064 51.479 to 60.817
15 58.030 2.157 53.151 to 62.909
16 59.912 2.257 54.807 to 65.017
Note d2 = 2 e2l(t — 2), with t — 2 degrees of freedom.

The interval is
c + c
fe t-2êT^T t-2êT)
and
square error. Thus if we use this as we can derive prediction

intervals for ocT as we did in Chapter 4. However, this interval gets
wider as T increases. This is illustrated in Table 5.3. In deriving our
forecast we have chosen the least squares line and extrapolated. It is
possible to show that this approach leads to the forecast with the
smallest mean square error. This is not true for all models, but it is true
with models that are linear in their parameters.
5.2 The local linear trend model
In the same way as the permanency of a mean may be open to doubt

so, too, may the permanency of a trend line. So we consider in this
section the forecasting of data which is not adequately described by a
global trend model but for which trend lines provide reasonable models
for each locality of time. Table 5.4 gives a set of sales data for which a
local trend model seems appropriate. Figure 5.3 shows a plot of this
data. As in section 4.2 it might be advisable to investigate more
64
Table 5.4 Data showing local linear trends
t xt t xt t xt t xt t xt
1 42.8 11 40.8 21 38.8 31 38.1 41 35.8

2 42.6 12 40.5 22 38.8 32 37.8
3 42.5 13 40.2 23 38.8 33 38.1
4 42.4 14 39.9 24 38.6 34 37.7
5 42.1 15 39.6 25 38.4 35 37.5
6 41.8 16 39.5 26 38.4 36 37.2
7 41.5 17 39.5 27 38.4 37 37.0
8 41.3 18 39.2 28 38.5 38 36.8
9 41.1 19 39.1 29 38.4 39 36.5
10 40.9 20 38.9 30 38.2 40 36.2
sophisticated models, but often a local linear trend will provide a simple
and robust model as the basis for forecasting. The terms in the more
sophisticated model might well describe the past more accurately than
our local trend model, but we may have doubts about their continuing
stability in the future. It will often seem in practice that the linear
trend is the only stable aspect of the situation that one can reasonably
expect to continue into the future, at least for a short time. This
distinction in attitude to past and future must always be borne in mind
by the forecaster. What he regards as a good model will often be
different from that, of the person whose only aim is to obtain a good
65
model for a given set of data from the past. We shall consider in this
section three approaches to forecasting with this model.
5.2.1 Holt’s method

In discussing exponential smoothing the estimate lxt' of the
underlying mean was expressed as a recurrence relation. There we used
the notation xt for the estimate of p, which was also the forecast of the
future observation xt+h • To emphasize that the equation did refer to
estimates, jOq, of the current mean, we shall rewrite the recurrence
relation as
fh = (1 ~a)xt + afrt- I
This is essentially a weighted mean, with weights (1 — a) and a of the
latest data xt, which estimates the current mean and the old estimate
jj.t_ j. In Holt’s method (see Coutie and coauthors, 1966; Holt, 1957)
we seek to produce corresponding recurrence relations for the estimates
of the two parameters in the trend model. However, rather than refer to
the parameter a, which was the mean of the model at time zero, we
rewrite the model so that a. is replaced by pf, the mean of the process at
the present time t. Thus at time t — r the model is
Xf— r ~ flf fir + — r
where r is time-measured into the past. By its definition and the fact
that the slope is fi, it is clear from Figure 5.4(a) that adjacent values of
fi are related by
= lit-1
Hence an estimate of \xt can be found from the two previous estimates
of j\it_ i and fi by simply adding fit_ ! to j£tf_ i, see Figure 5.4(b). This
gives the estimate of the mean at time t based, on past data. The latest
observation xt also provides an estimate of pf. Hence if we use the ideas
in the exponential smoothing recurrence equation and make the new
estimate jlt a weighted mean of the latest data and the estimate based
on using the previous estimates, we arrive at
— +
ft = (1 a)xt + I fit- I )
To make use of this equation, we need a way of calculating the
sequence of estimates fit_ l. This may be done by applying the idea
used in getting the above reccurrence relation to finding a recurrence
relation for the fi^ values. The new value fit will be a weighted mean of
the old estimate fit_ { and an estimate based on the most recent data.
To estimate the slope using the most recent data, we might use
xt — jcf_ i, since from the model this can be written as
xt — xt_ ! = (iit + et) — (iit+1 — fi + et_ !)
= fi + et-et_1
66
(a ) Model structure (b) Two estimates of /x/
(c) Two estimates of /?/
A A
H-t - H-t-1
Figure 5.4 The basis of Holt’s method
Thus xt — xt_ ! has expectation /3 and thus provides an estimate of (3.

Unfortunately, because it involves both et and et_ x, it has a large
variance and so the estimates of f3 would not be very stable.
Figure 5.4(c) suggests a better way of estimating. This is to use the
relation
0 = M/ — Mf- l
By replacing the unknown parameters jut and fJLt- I by our estimates of
them, we see that gives an estimate of (3 and includes
implicitly the influence of the latest observation xt. It is, however, not
as strongly influenced by xt as the previously suggested estimate. The
method of updating the estimate of slope is thus based on the
recurrence relation
Note that we do not have to use the same weighting constant as for jCq,
so a new constant b is introduced.
The practical application of these formula requires the estimates of (3
and ju to be updated alternately using the recurrence formulae above,
each updated estimate being used in the next calculation of the other
parameter. The calculation is illustrated in Table 5.5. We now have a
way of finding estimates of the slope and current mean at the time of
67
Table 5.5 Example of Holt’s method
(a) Formulae:
+
£* = (!“ a)xt + a
(Pt-1 fit-1)
+
fit = (1 - &)(£* ~ £t-i) b$t_i
^r+2 =
fit +
%fit
a = 0.8 6 = 0.9
Initial values assumed pi = 42.9, (3 = —0.15
(b) Data from Table 5.4
0.2 0.8 0.1 0.9
xt P-t-1 +
fit-i Pt Pt-fit-i 6t-1 ft x
t+ 2
— —
42.900 — —
-0.15 —
42.8 42.75 42.760 -0.140 -6.150 -0.149 42.46

42.6 42.611 42.609 -0.151 -0.149 -0.149 42.311
42.5 42.459 42.468 -0.141 -0.149 -0.148 42.172
42.4 42.320 42.336 -0.132 -0.148 -0.146 42.044
42.1 42.190 42.172 -0.164 -0.146 -0.148 41.876
41.8 42.024 41.979 -0.193 -0.148 -0.152 41.675
41.5 41.826 41.761 -0.218 -0.152 -0.159 41.443
the most recent observation. The estimated equation of the line at time
t is
A
xt_ r = fit fitr (r = 0, 1, . . . , t — 1)

A forecast of xt+h is thus provided by
xt ,h = Ar + ft*
obtained by extrapolating along the estimated line as in Figure 5.4(d).

As in the constant mean model, the independence condition on the e
values means that we can only forecast the deterministic part of the
situation and not the future random variable et+h.
To start forecasting using recurrence relations, we need good starting
values p0 and /30, see Figure 5.5. These values are best based on an
initial set of data. From this set of data an initial line can be obtained
A
by any of the methods of sections 5.1. j30 is then taken as the slope of
this line and p0 as its value at the time one unit before the initial
observation is obtained.
If there is an initial set of data available, it is wise to use it to obtain
some idea as to the best values for the forecasting parameters a and b.
For example, if we take the mean square error as the criterion, then we
68
Figure 5.5 The effects of initial conditions
can evaluate the mean square error for a range of different values of a
and b. Since all methods produce poor forecasts when there is very
little data, the forecasts obtained over the first few observations should
be ignored in calculating the mean square error. A discussion of
methods of choosing parameter values is given in section 15.3.
In Chapter 4 we noted that the recurrence form of forecasting
formula could also be written in terms of the forecast errors. This can
also be done here. The one-step-ahead forecast error is
A
et = xt — {Ut- I +&_.)
Substituting the value of xt given by this definition in the recurrence
relation for jdt gives
Mr = (1 — a) {e, + (fit- i + ft- i)} + a(Ar- i + Pt- 1)

Hence
fit = ixt_ i + Pt-i + (1 ~ a)et

which is the obvious extension to a trend line of the error correction
form of Chapter 4. If we substitute^ the expression of fit — fit_ j from
this into the recurrence relation for /3t, we get
0t = (1 - b) $t- i + (1 - a)et} + b$t_ !

Hence
Pt = Pt-i +cet
69
where c = (1 — fr)(l — a). Thus we can update both the mean and slope
estimates by adding simple multiples of et.
5.2.2 Discounted least squares

In the discussion on the least squares method on section 5.1.2 it was
stated that it was reasonable to choose the parameters in the model to
minimize the sum of the squares of the distances of the observations
from the line that we propose to fit to the data. This sum of squares we
denoted by
S= Z 6/2
t= I
where the e values were called the residuals. In doing this we attach as
much importance to residual ex, for our very first observation, as to the
residual eu for our latest observation. When we are dealing with data
believed to come from a local model and our aim is to get a good fit to
the recent data for forecasting purposes, then this equality of
importance may not be very sensible. It may be more reasonable to
introduce a weighting, or discounting, into the sum of residuals squared
in exactly the same way that we introduced it into the ordinary mean
in justifying use of the weighted mean of exact exponential smoothing.
Various types of weighting or discounting could be introduced but, as
in the constant mean model, it is found that the use of exponential
weights lead to particularly simple calculations. We thus replace the
sum of residuals squared by a weighted sum in which e]_ r has a
weight of ar (0 < a < 1) associated with it. The most recent observation
gives a residual squared with weight one. The far-past residuals, having
small values of ar (for a < 1), contribute little to the total. As the
magnitude of a controls the extent to which the past is discounted, it is
often called the ‘discounting factor’. Thus our least squares criterion is
replaced by
S=Vae _
r= 0
r 2
t r
which is the discounted least squares criterion (see Gilchrist, 1967). We

can now substitute the value of et_ r given by our line, again taking the
present as origin for r, i.e.
€f— y ~ Xt — r frt
A
We now choose the values of and to minimize S. This is done in

detail in Appendix D. The solution can be put in the form of normal
70
equations, as in section 5.1, the new normal equations being
t- 1 A t- I t-I
r r
fit 1 a -pt 2 m = 2
r=0 r=0 r=0
fit 2 rar — fit X r2 ar = 2 ayrxtir

r—0 r=0 r=0
It will be seen that the form of these equations is the same as that for
the normal equations for the undiscounted case — which it must be, for
this corresponds to putting a - 1. The sign of fit has changed since we
are now measuring the time, r, as positive into the past. The other
difference is that all the sums are now ‘discounted’ sums:
lx, is now T.arxt_ r,
t t- I
f = 2 1 is now 2 ar,
1 o
and so on for each term. To get the full value of introducing the
weights a, we require evenly spaced data, so we shall assume in what
follows that observations occur evenly at times 1, 2, 3, ... . The
occurrence of uneven intervals can be dealt with fairly easily once the
principles have been grasped.
The forecast of xt+h is, as before,
X, ,h = M, + isth
The solution of this forecasting problem looks computationally very

difficult, but all the discounted sums required can be calculated very
simply by using recurrence relations.
Table 5.6 shows the recurrence relations that can be used for
updating the various terms in the normal equations and their
application to a set of data. It also indicates the limiting values of the
various terms which approach constant values as the amount of data
increases.
The calculations used in this method are clearly more involved than
those of the previous method discussed. In many circumstances,
however, there are advantages in using this method. It does not require
the specification of initial conditions and thus can be used with some
confidence when only small amounts of data are available.
If we set a = 1, the above normal equations become identical with
those of the ordinary least squares method. If we have a local constant
mean model, corresponding to fi =,0, we obtain the exact exponentially
weighted moving average of the previous chapter as the discounted least
squares estimate of ju.
71
Table 5.6 Example of discounted least squares
(a) Recurrence relations for the normal equations
Initial
Sum Symbol Recurrence relation value
t-1
Wt Wt = 1 +aWt_x Wx = 1
r-0
t-1
1
2 mr
+
Ax =0
0
ii
A,
'■+
1
r-0
t-1
2 r
2 r a B, Bt =At + a(At_l + Bt_i) Bi =0
r=0
t-1
r
2 a xt_r Yt Yt
= x
t
+ a
Yf-i Yi
r—0
f-1
r
2 ra xt_r z, Zt = a
(Yf-i
+
Zt_ i) Zj =0
r=0
A A
Normal equations Wtidt —At@t = Yt

=
A-tfi-t Btfit Zf
Solutions of normal equations
=
Bt Ct(BtYt — AtZt)
Pt = Ct(AtYt-WtZt)
where
Ct = 1 l(WtBt —A]) — note this is independent of data
00
Limiting values as t -> a = 0.8 a = 0.9
W- 1/(1 —a) 5 10
A = a/(l — a)2 20 90
B = a(l + a)/(l — «)3 180 1440
1/C = a/(l - a)4 500 6300
c 0.002 0.000158
(b) Example, using data from Table 5.4

a = 0.8
At B, xt Yt Ct Bt pt *7+2
1.000 0.0 0.0 42.8 42.8 0.0 — — — —
1.800 0.800 0.800 42.6 76.84 34.24 1.25 42.6 -0.2 42.200
2.440 2.080 3.328 42.5 103.972 88.864 0.2626 42.487 -0.149 42.189
2.952 3.616 7.942 42.4 125.578 154.269 0.09644 42.386 -0.127 42.132
3.362 5.254 14.501 42.1 142.562 223.877 0.04730 42.146 -0.173 41.800
3.689 6.893 22.697 41.8 155.850 293.151 0.02761 41.874 -0.198 41.478
3.951 8.466 32.138 41.5 166.180 359.201 0.01808 41.578 -0.223 41.132
72
5.2.3 Double exponential smoothing
In section 4.2 we made use of the recurrence formula for the

exponentially weighted moving average. This was
5c t = (1 — a)xt + axt_l
If we carry out ordinary exponential smoothing on the data of

Table 5.4, it will be seen that the value of 5c t lags behind the trend (see
Table 5.7). To find the expected lag, /, we substitute the expected value
of xt_r, for the trend model, in the exact form of exponential
smoothing. Thus
E(5ct) = Xal'(nt-l3r)IZar
= 1ut-P(2rarIXar)
= M —f a/(l — «)}/3
for large f, using the limiting results given in Table 5.5. Thus the lag is
{a/(l — a)}p.
A reasonable way to obtain an estimate of the current mean is to
adjust x t by adding an estimate of this lag. One way of doing this is to
make use of what has been termed double exponential smoothing (see
Brown, 1959, 1963). Since the exponentially weighted moving average
has the effect of smoothing out much of the fluctuation in the data x x,
x2, x3, . . . , xti the sequence 5c l, 5c 2, x3, . . . 5ct is said to be the result
of exponentially smoothing the first sequence. Suppose that we now
apply the formula of exponential smoothing to
This we could denote by
x/2 )
= (1 — a)5ct + a5c)2_\
The new sequence 5c{2 \ 5c^2 \ 5c^2\ . . . , x[2^ is said to have been
obtained by the process of ‘double exponential smoothing’ or second-
order exponential smoothing of the data. If xt exactly follows the trend
line, then xt lies on a trend line a distance l below, when p is positive.
Similarly x}2) will be a distance / below 5ct. Thus we would have
xt = 5ct + l
X — X^ ^ ^ -4-7
By subtracting we get
xt = 2 5ct — 5c
We have here assumed a perfect line. Where xt is just following a local

trend line, then we use the right-hand side as a reasonable estimate of
the current mean fjLt. Hence
At = 2*r —x\x)
73
To estimate the slope, we use the facts that for the exact trend
l ={a/( 1 — a)}/3
and
1 — x
l
-y* ( 2 )
*/V f
Hence a sensible estimate of (3 at time t is given by
1ft = -—- (xt - x{2))

a
These are obviously much easier equations to use than the normal
equations, though it must be remembered that these are only
approximations valid for large t since the lag is only l for large t. For
small t these equations can still be used, almost as a separate method,
provided we have enough prior information to enable us to put in
sensible starting values for x0 and Brown, who developed this
method, shows that, if we have p0 and ft as initial estimates of the
mean and slope at the time of the beginning of the data, then
a
*o = Mo ~ “ ft
1—a
and
xtf) = Mo - : ft
1 —a
give the correct starting values to use. If we can at best make rough
guesses for p0 and ft, it is advisable to use the exact method as given
by the normal equations, at least to obtain the first few forecasts.
Table 5.7 provides an example of the method.
As with Holt’s method, we may express the forecasting formulae in
terms of the one-step-ahead forecasting errors. The derivation of these
formulae is an exercise in algebra which we shall leave to the more
mathematically inclined readers. The final equations in error correction
form are
A; = At- i +ft_ i + (1 ~a2)et
ft = ft- I + (1 —a)2e,
A comparison of the above error correction forms with those of Holt’s
method shows that the double smoothing method is just a special case
of Holt’s method. In fact, if we introduce subscripts h to denote the
parameters in Holt’s method, then the two sets of equations are
identical if we write
ah = a2
74
Table 5.7 Example of double exponential smoothing
(a) Formulae
xt = (1 — a)xt + axt_i
3cf> = (1 — a)xt + ax^
jlt = 2xt — x}2}
Pi-
a
=
a = 0.8,/JQ 42.9, = —0.15, assumed initial values giving
x0 = 43.5, X<2) = 44.1
(b) Data from Table 5.4
xt xt fit ft Xf+2
43.5 44.1 42.9 -0.15 —
42.8 43.32 43.944 42.696 -0.156 42.384

42.6 43.176 43.790 42.562 -0.153 42.256
42.5 43.041 43.641 42.441 -0.150 42.141
42.4 42.913 43.495 42.331 -0.145 42.041
42.1 42.750 43.346 42.154 -0.149 41.856
41.8 42.560 43.189 41.931 -0.157 41.617
41.5 42.348 43.021 41.675 -0.168 41.339
and
ch = (1 — a)2
Thus if in Holt’s method the parameters are related by
ch = (1 — sfa^)2
the method is identical to the double smoothing method.

Double exponential smoothing is also related to the method of
section 5.2.2, since it can be shown that for large t the solutions of the
normal equations are equivalent to the estimates of \ut and (3t given by
the use of double exponential smoothing.
5.2.4 Conclusions
We have now considered three methods of estimating the parameters

and forecasting for a local linear trend model. Table 5.8 gives a
summary showing the relations between them. The practical decision as
to which method to use requires the choice of the best forecasting
parameters (a or a and b), which is discussed in section 15.3, and the
comparison of the methods on real or suitably simulated data.
75
Table 5.8 Summary of methods of section 5.2
Discounted least Double exponential

Holt’s method squares (DLS) smoothing
Basis of derivation Construction of Estimation of Construction of

recurrence model parameters unbiased estimators
relations based on using method of out of single and
model structure DLS double exponentially
smoothed data,
assuming t large
Number of
forecasting
parameters 2 1 1
Initial estimates of
lit and Pt required Yes No Yes
Relations between A special case of

methods Holt’s method
Equivalent to DLS
for large t
The reason for discussing all three methods at some length was not
just to cover the main methods. The three methods we have discussed
illustrate three main approaches to forecasting local models, which are
(a) to use discounted versions of classical statistical methods

(section 5.2.1).
(b) to develop intuitively recurrence relations for the parameters
(section 5.2.2).
(c) to use a simple method and then adjust for the bias, or lag
(section 5.2.3).
We will come across all these methods again in later chapters.
5.3 Polynomial models
An obvious extension of the trend model is to move to polynomial

models involving t2, t3 , etc., to describe more elaborate movements in
the data. All the methods described in this chapter have natural
extensions that provide forecasts for such situations.
A word of caution is, however, advisable in relation to using
polynomial models. As a rule quadratic trends and higher order
polynomials should only be used if there is some logical basis for such a
structure. In general the magnitudes of errors both from incorrect
choice of model and from changes in structure in time increase with the
76
complexity of the model used. Further, the more parameters used the
more observations are needed to obtain forecasts. Even with global
models a lot of data are required to give a reasonable error variance.
Cowden (1963) tabulates the type of information required to assess the
error variances. By way of brief illustration, if, as usual, the e values are
independent with variance o2, then for, say, eleven observations the
one-step-ahead forecast error variance is 1.647 o2 for a linear trend
model, 2.098 o2 for a quadratic model, 2.771 a2 for a cubic model and
3.908 o2 for a quartic model. Clearly, in a less stable situation with a
model fitted by, say, discounted least squares the error variances will be
even higher in practice.
References
Brenner, J. L., D’Esopo, D. A., and Fowler, A. G. (1968). Difference equations in
forecasting formulae. Man. Sci., 15, No. 3.
Brown, R. G. (1959). Statistical Forecasting for Inventory Control McGraw-Hill,
New York.
Brown, R. G. (1963). Smoothing; Forecasting and Prediction of Discrete Time
Series. Prentice-Hall, Englewood Cliffs, New Jersey.
Coutie, G. A., Davis, 0. L., Hassall, C. H., Miller, D. W. G. P., and Morrell, A. J. H.
(1966). Short Term Forecasting. ICI Monograph No. 2. Oliver and Boyd.,
Edinburgh.
Cowden, D. J. (1963). The perils of polynomials. Man. Sci., 9, No. 4, 546—550.
Croxton, F. E., Cowden, D. J., and Klein, S. (1968). Applied General Statistics.
Pitman.
Gilchrist, W. G. (1967). Methods of estimation involving discounting. J. Roy.
Statist. Soc. (B), 29, 355—369.
Holt, C. C. (1957). Forecasting Seasonals and Trends by Exponentially Weighted
Moving Averages. Carnegie Institute of Technology, Pittsburgh, Pennsylvania.
Chapter 6
Regression models
6.1 Introduction
In the previous chapters we have considered methods of forecasting

that only make use of past values of the variable being forecast. Clearly,
this approach is limited as it is reasonable that to obtain the best
forecasts we may need to make use of information from a wide variety
of sources. In particular, we need to make use of the information
contained in other variables that are related to the one being forecast.
For example, in forecasting the sales of one particular product we may
relate the sales S to the wealth in the consumer’s pocket, as measured by
some income variable /, the price of the product P, the expenditure on
advertising A, and probably other variables as well. We would thus be
looking for a model of the general form
S = f(I9P,A, . . .)
where/( ) is some function to be found. Our ability to construct such
models that provide good forecasts is a measure of our real under-
standing of the situation. Clearly, in constructing models for this type
of forecasting a good knowledge of the underlying situation is essential.
There are a number of books that deal particularly with this area of
forecasting, e.g. those by Robinson (1971) and Spencer, Clark and
Hoguet (1961).
Within the scope of this book we are concerned only with some of
the statistical aspects of the use of such models. There are an infinity of
possible models that might be used, so we must clearly limit our
discussion to a simple type as an introduction to the subject. In this
chapter we will consider the problem of forecasting with the most
commonly used model, namely that based on the use of linear
relationships. By way of example, suppose that in the situation
described above we finally obtained the model
St = 127.12 + 2.34 It - 0.32 Pt + 0.65 At_ j + 6.82 t + et
77
78
This model is linear in its -parameters. It will be seen that St depends

not only on the current values of / and P but on the previous value of
A and on an underlying trend represented by the 6.82 t. As usual, the
random variable et represents the failure of the world to behave in the
nice regular fashion that would make forecasting so much easier. To
consider the forecasting use of such models, let us limit the considera-
tion to the simplest case where the model relates a variable yt to just
one other variable, xt. The linear model would thus be
yt =a + (3xt + et
If we can use the data to estimate the unknown parameters, a and j3,
given estimators a and 6, then we have as a fitted model the line
y = a + bx
These forms of model are called linear regression models and we talk
about fitting a regression line of y on x. If we wish to forecast yt+h , the
natural thing to do is to use
—
yt+h C + bXf+h
This is fine, provided that we know the future value xt+h. There are
situations where this is the case, as, for example, where xt+h is the
number of selling points of a product, in which case xt+h is a value
under our control. In general, however, xt + h may be as much a
quantity requiring forecasting as yt+h. If xt+h is a forecast of xt+hi
then
yt + h =a + bxt+h
will provide a forecast of y t+h _ This forecast may be no better than that
obtained by looking at past values of yt and using, say, a local trend
model. It does, however, provide another forecast based on some
additional information. This may be usefully combined with the trend
forecast using the methods to be discussed in Chapter 17.
It is sometimes the case that past experience shows that xt+h leads
to a better forecast of the y variable than we could obtain from past
values of the y variable. In this case, using the regression model provides
a good way of forecasting yt+h. As an alternative procedure worth
considering, we could examine as a basic model
yt = a + fixt + et
where xt is the forecast of xt. We would then fit this to past values of y
and x to obtain a fitted line
y =a + b x
and forecast yt+h by

79
yt+h = a + b'xt+h
In a different type of situation a change in a variable x may not
produce an effect on y for some time, h. For example, sales of spares y
are likely to lag behind sales x of the item for which they are spares. In
such a situation we might have a regression model of the form:
yt = a + pxt-h +
where xt_ h is referred to as a lagged variable. We can now forecast yt+h
by fitting and extrapolating to get
A rr , 7 n
yt+h = a + b xt
where xt is the latest known value of x.
Clearly, it is more common to have the variable y depending on
more than one other variable, e.g. ice cream sales may depend on both
temperature and sunshine. If we denote these variables by xl t,
x2 • • • , xkt, then we may write our general model as
k
yt = 2 ft*,. ( + et
i=0
The constant term is allowed for by simply letting x0 t = 1 for all t;

then p0 is the constant. We will use this form as a general notation; the
xi>t may represent an observation at time t, a lagged observation at
some previous time, a forecast of a variable at time t or possibly a
function of time. These xi t values, which we will term the regressor
variables, are regarded as a set of given constant values. The variable yt,
the dependent variable, is thus composed of a deterministic part,
EPiXift9 and an added random variation, et.
Before turning to some of the details of fitting and forecasting
regression models, it is important to emphasize that our ultimate
concern is in forecasting. Much of the statistics concerned with
regression deals with measuring how good the model was and how well
past y were explained by the right-hand side of the regression equation.
It can easily happen that a very good model, looked at in purely
statistical terms, can lead to very poor forecasts because of problems of
reliability of data and of inability to get good forecasts of regressor
variables. Our ultimate criteria of success cannot be the various
measures of fit described in textbooks dealing with regression; our
criteria must be criteria such as the mean square forecasting error that
refer directly to the forecasting use that we make of the regression
model.
6.2 Fitting regression models
We have discussed the form of regression models used for forecasting.

We now turn to examine briefly the question of fitting these models.
80
There is a great deal of literature on this topic (see, for example, Draper
and Smith, 1966). In nearly all this literature the models are regarded as
being, in theory, global models, though in practice the users of the
models often ignore data from the far past as being irrelevant. From the
forecasting viewpoint it seems more reasonable to assume that the
models are best treated as local models. Thus rather than use least
squares we would use discounted least squares with a discounting factor
ar. If it so happens that we are wrong here, then in choosing a to get the
best forecasts we find, approximately, that a- 1. We would thus obtain
the same forecasts as the user of ordinary least squares fitting.
To indicate the type of result we obtain by applying discounted least
squares, consider the data in terms of the fitted model
k
yt = 2 b,xut + e,
i= 0
where et is the error in the fitted regression, the residual. The regression
coefficients bt are the discounted least squares coefficients when they
are chosen to minimize
s = tx\ri?_r
r= 0
Substituting for e and minimizing with respect to b0, putting x0 t = 1,

leads to
t- 1 / k \
^ ^ (yt — r ^ t—r] ~~ 9
r~ 0 \ f=0 /
Minimizing with respect to any bt gives

1
t- / k >
—
2 ci Xj ^_r[y^_r 2 bfXf,t — j I 0
r=0 \ i= 0 /
(see Appendix D). If we denote,

t- I
2 arzt_rvt_r
r= 0
by S(z, v), the above equations giving the b values, are

b0S( 1,1) + blS{lix1) + ... + bkS( 1, xk) = S{ 1, y)
+
b0S(xi > 1) blS(x1, Xj) + . . . + bkS(xj, xk) = S(xx, y)
b0S(xk, 1) + b1S(xk, Xj) + . . . + bkS(xk, xk) = S{xk,y)
These equations, the normal equations, can be compared with the

81
equation of the fitted regression

b0 + bx xx + . . . + bk xk = y
Thus the k + 1 normal equations are obtained by multiplying this
equation by a\ arx1, . . . ,arxk, respectively, and summing each over
time. Solving these equations gives the coefficients b0, . . . , bk. Table
6.1 gives an example of such a calculation. In (a) the ordinary least
squares line is fitted. Clearly, in the calculation the order in which the
observations occurred is irrelevant. In the normal equations S(u, v)
represents the sum of products of u and v. To calculate the normal
equations in the discounted case, we note that if St(u,v) denotes the
discounted sum at time t then
St(u,v) = utvt + aSt~ ! (u,v)
where S1(u,v) = ul vx. Therefore we can use the usual type of
recurrence relation to calculate these quantities and thus write down
the normal equations at any time t. The normal equations of Table
6.1(b) are those at time 12. Figure 6.1 shows the data and the fitted
lines for a = 1.0, a = 0.95 and a = 0.90. It will be seen that as a
decreases the line moves closer to the most recent data points, as one
would wish it to for forecasting purposes.
Almost all computers and the more sophisticated desk-top calcula-
tors provide programmes that solve the normal equations and estimate
the regression coefficients b j using ordinary least squares. These can be
used to obtain the discounted least squares solutions by discounting all
the regressor variables xi t_r (including x0>f_r) by multiplying each by
the value of (\/a)r. The logic of this is shown by the following
reformulation of the discounted least squares criterion:
Y.are]_r = t\.{sja)ryt_r-bâ)rx0it_r - .. . — bk(yja)r xkjt_r]2
In using forecasts from a regression model we will wish to develop

some idea of the error variance. Let us look briefly at this for the
simplest model
yt = fixt + et
If xt+h is a known quantity, then
yt+ H ~ bXf+k
and
Var(ef+Aj) = Var(yf+/Z — yt+h)
= Var(j3xf+fc + et+h — bxt+h)
= Var’{b)x2t+h + o2
Notice that in this simple case the variance of the error increases as the
82
Table'6.1 Fitting a regression line

(a) Classical calculation
2 2
t y X xy xz y
1 1.2 28.3 33.96 800.89 1.44

2 1.3 26.8 34.84 718.24 1.69
3 1.3 28.7 37.31 823.69 1.69
4 1.6 35.1 56.16 1232.01 2.56
5 1.6 37.0 59.2 1369.00 2.56
6 1.4 30.5 42.7 930.25 1.96
7 1.5 32.5 48.75 1056.25 2.25
8 1.6 35.7 57.12 1274.49 2.56
9 1.5 34.4 51.6 1183.36 2.25
10 1.7 37.3 63.41 1391.29 2.89
11 2.0 39.2 78.4 1536.64 4.00
12 2.1 43.2 90.72 1866.24 4.41
Totals 18.8 408.7 654.17 14182.35 30.26
Normal equations, a = 1
M(1,1) + bxS( l,x) = s(i,y) = 12 b0 + 408.7 bi = 18.8

&0S(1,*) + biS(x, x) = S(x, y) = 408.76 + 14182.35 bl = 654.17
bo = -0.232
bi = 0.053
(b) Discounted regression, based on table of (a), using recurrence form

st( u, v) = uv + aSt_ i (u, v)
t Zaryt_r li,arxt_r 2(/xf_ ry t—r larx2t_r Xary2t_r IV'
1 1.2 28.3 33.96 800.89 1.44 1

2 2.44 53.68 67.10 1479.09 3.06 1.95
3 3.62 79.70 101.06 2228.83 4.60 2.85
4 5.04 110.82 152.16 3349.40 6.93 3.71
5 6.38 142.28 203.75 4550.93 9.14 4.52
6 7.47 165.66 236.27 5253.63 10.64 5.30
7 8.59 189.88 273.20 6047.20 12.36 6.03
8 9.76 216.08 316.66 7019.33 14.30 6.73
9 10.77 239.68 352.43 7851.72 15.84 7.39
10 11.94 265.00 398.22 8850.43 17.94 8.03
11 13.34 290.95 456.71 9944.55 21.04 8.62
12 14.77 319.60 524.59 11313.56 24.40 9.19
Normal equations, a = 0.95 and t = 12

9.19 b0 + 319.60 bx = 14.77
319.60 b0 + 11313.56 bx = 524.59
b0 = -0.284
bi = 0.054
83
Figure 6.1 Plot of the data and regression lines of Table 6.1
square of the distance of xt+h from the origin, which acts as a pivot. In
the general case the exponentially weighted mean of the x values acts as
the pivot. It is the uncertainty in the estimate of slope that generates
the element Var(b)x2+h in the error variance.
If xt+h has to be forecast by oct+h , then
Var(er+„ ) = Var(pxt+/l + et+h - bxt+h)
= Var(bx,+ „) + a]
= E(b2)E(x2+h)—p2xj+h + o2e
= (Var(6) + p2 }{ Vax(xt+h ) + xj+h} — p2x2t+h + o2e
= Var (b)xj+h + o2 + Var(xf+A X Var(6) + p2}
assuming that b and xt+h are independent and xt+h is an unbiased
forecast of xt+h. Thus the expected mean square error is increased by
an additional amount depending not only on Var(xf+^) but also on
Var(6) and (32. As usual, we need to investigate the magnitude of the
mean square error directly via a study of the observed errors when the
method is applied to past data. Theil (1966) gives a general study of
forecast errors in regression models.
6.3 Selecting the variables
In choosing the variables to use in regression for forecasting one must

of necessity start by using one’s knowledge of the situation to pick
84
likely candidates. Very often, however, there are many possible

candidates and it is a real problem to choose a small number to use for
one’s actual forecasting exercise. The need for a small number of
regressor variables is not at first obvious. If we only have a few
observations and a comparable number of variables, then the fitted
regression could be made to fit very closely to the data. However, this
would be picking up the peculiarities of the individual observations and
the variances of the b values would be high. The variance of yt+h would
therefore be unacceptably large. To get reasonable forecasts we
therefore need many more observations than variables. This in practice,
where data is often short, puts a constraint on the number of variables
used. A further aspect is that in many practical situations most of the
variability in y can be explained in terms of a few of the variables.
Though the fit can always be improved by introducing new variables,
the amount of the improvement may be small and the quality of the
forecasts may well deteriorate.
The problem thus arises as to how to choose a small number of
regressor variables from the original set chosen. There are a large
number of different ways of doing this. e.g. see Allen (1971) and Beale
(1970).
In common practice most methods assume a global model and are
based on:
(a) repeated examinations of the correlation of the fitted regression
and the dependent variable y;
(b) studying the changes in this produced by adding or removing the
various regressor variables to and from the model;
(c) studying the significance of the regression coefficients obtained.
They are thus based on the use of statistics such as the correlation
coefficient and t statistic.
The amount of calculation involved in these methods is very large
and a computer is really essential. All major computers have pro-
grammes available that can be used in the selection of variables.
From the forecasting viewpoint the assumption of a global model as
the basis for selecting variables is not so reasonable. It is clearly
advisable to select the variables that provide the best explanation of the
most recent data rather than the averaged-out sort of selection made by
the above methods. If we admit that our real data may not be from a
stable, global, situation, but rather that a local approximate model-
building approach is more appropriate, then two things follow. Firstly,
the assumptions underlying the above methods will not be valid.
Secondly, it would intuitively appear both reasonable and possible to
replace the correlation coefficients, t, statistics, etc., of the above
methods by modified versions based on discounted statistics. These can
then be used in modified versions of the above methods. A detailed
85
discussion of such an approach is given by Singleton (1971). By way of

a simple illustration, consider the multiple correlation coefficient R,
defined by
R2 =1- 2 if
i= 1
/i
/ i= 1
(y,-y)2
where
y= 2 yy/t
Z= 1
This is the simplest measure of the ability of our model to ‘explain’ the
data. If our model fitted perfectly, the e values, the residuals, would all
be zero so R2 would be one. If it did not really ‘explain’ the data at all,
the variability of the e values would be much the same as that of the y
values. In this case Ze2 would be approximately the same as
2(yz ~y)2> so that R2 would be approximately zero. Most of the
methods above make some use of the magnitude of R2 in selecting the
regressor variables. If we now say that interest is concentrated on the
most recent information, then it is reasonable to introduce discounting
into the above statistics. Thus if our usual exponential weights are used,
we would replace
t
e
^ i
i= 0
t
2 are2t- r
r= 0
and
t
2 (y; - y?
2=0
by
2 ar(yt_r —y)2,
r= 0
where y may be the ordinary mean or the exponentially weighted

mean. We would thus define a discounted multiple regression co-
efficient by
*3=1- 2 are2_ r / 2 ar(yt_r~y)2

r-0 / r-0
86
Table 6.2 Some illustrative values

of R2d
Discounting Variables Variables

factor (3, 5, 6) (4, 5, 6)
a 4 4
1.00 0.83 0.81
0.99 0.90 0.89
0.98 0.94 0.94
0.97 0.96 0.96
0.96 0.96 0.97
The properties of Rd of relevance to the selection of variables are

essentially the same as those of R2, so it can be used to form the basis
of methods for the selection of variables that are most important locally,
which is the object in forecasting situations.
Table 6.2 shows the values of Rd from an exercise in which there
were five regressor variables (labelled 2, 3, . . . , 6). The values of Rd are
used to compare the effectiveness of the sets (3,5,6) and (4,5,6). It is
seen that for a = 1.00, i.e. for the usual multiple correlation coefficient,
the first set is better than the second. However, if we put more
emphasis on recent data and set a = 0.96, then we begin to get the
opposite picture.
We have referred to the use of lagged variables as regressors. In
practice one is not often sure as to the length of the lag. In initial
explorations one therefore needs to put in a range of different lags as
possible regressor variables. It is then up to the regression program to
select the most appropriate ones. The selection of variables in this
fashion gives the selection of regressors for obtaining the best fit or the
best forecasts over past data. It does not necessarily pick the variables
that are clear causal factors in determining the values of the
independent variable. It is then up to the forecaster to decide whether
the results obtained are chance results, depending on the chance form
of the data used, or whether they represent the real effects necessary to
obtain good forecasts in the future. To do this the forecaster, as always,
needs to know a great deal about the variables in the forecasting
situation. The use of the programs for selecting variables is thus part
of the preliminary exploration of the data, and the forecaster should
use a wide range of regressor variables in this exploration.
It may be helpful at this stage to list some of the types of variables
that have proved useful in forecasting situations:
(a) Forecasts of a variable within a firm, say its sales, often use various
measures derived at a Government level. As indicated before, these
might be actual values, forecast values or lagged values. A list of
87
sources of American economic and other variables of use in

forecasting are given in Woy (1965). A large number of business
and trade journals also publish relevant data of value for particular
industries.
(b) It is often of value to include functions of time in the regression
model. Thus we might have
=
yt (30 +
Pi t + p2x2,t + ef
A seasonal variation can be modelled, using weekly data, by using
harmonic terms. For example,
= +
yt Po Pi cos 2nn/52 + P2 sin 27ra/52 + P3x3 >f + et
where n = 1, 2, . . . , 52 is the week number in the year.
(c) A rather similar device is sometimes used to deal with quarterly or
monthly seasonals or the effects on sales of holidays, etc. Here we
define a series of x variables that are zero everywhere except at the
month or holiday week in question. The value 1 is given at that
time. The term px is only non-zero at that time and then its
estimated value b. 1 gives the magnitude of the seasonal or holiday
effect. If there are monthly seasonals or a large number of holidays,
we have a corresponding number of parameters whose values
therefore cannot be estimated as precisely as when only a small
number were used. One way round this, as a fairly rough way out,
is to use past data to invent a function whose shape seems about
right for the relative magnitudes of the various effects. For
example, it might be that the holiday effect is: 1 in weeks 12, 19
and 35, 2 in weeks 3, 30 and 48, and zero elsewhere. The
coefficient, when estimated, will then give an improved estimate of
the level of the holiday effects, though obviously, in the example,
it would not alter the 2:1 ratio chosen rather arbitrarily. Thus if
the value of b was 0.7, as happened in a particular case, the
estimated holiday effects become 0.7 and 1.4 on the appropriate
weeks.
(d) Considerable thought must be given to the way in which any
regressor variable will influence the dependent variable. It is
sometimes the case that it is only the general level of a particular
regressor that is important. The random variation in this variable is
thus an irrelevance that will make our forecasts worse. An obvious
way of dealing with this is to try and smooth out this random
variable. The smoothed data would then be used as the regressor
variable. By way of a further example, in a sales forecasting
exercise, the effect of temperature was thought to be important.
However, on reflection an important feature was not the actual
temperature but whether that temperature was greater or less than
the norm expected by people at that time of the year. Thus two
88
series were used; one, Tt, was the seasonal normal temperature, the
other, ATt, was the deviation of the actual temperature from this
at time t. In the event it was found that Tt was effectively allowed
for by the seasonal terms in the model and ATt remained as a
highly significant term.
(e) Suppose that variables xx t and x2 ft are thought to be relevant for
forecasting the yt series. It does not necessarily follow that xx and
x2 should appear in the regression model as
= +
yt ,t &2X2 ,t +
et
It could be, for example, that the best model for forecasting is
yt =/3,x?>r +132 logx2,( + P3xt t, x2J + e,

In other words, the ways in which xXJ and x2J influence yt are
not necessarily simple or linear. Note, however, that the model is
still linear in the parameters and is still straightforward to fit.
Programmes to select variables can be used to look at the use of
functions of regressor variables, the functions being fed in as
though they were totally separate regressors. Care, as usual, must
be taken in the interpretation of the outcomes of such exercises,
but if the forecaster obtains a model that gives better forecasts
then the exercise is worthwhile.
6.4 Econometric models

As the title of this book is Statistical Forecasting the writer does not
wish to devote much space to economic forecasting. There are a
number of good books in this area (e.g. Theil, 1966) and to do the
topic any justice at all would require far more space than is available.
There are, however, many statistical aspects of economic forecasting
and, indeed, some books of econometrics could well have the same title
as this chapter. Again, however, much of the statistics involved arise out
of the nature of the economic theory used and will not be dealt with
here. The aim of this section is to show briefly how the subject links in
with what we have said so far. The reader may then know whether in
his problems he needs to move on to a study of econometric models.
The first and obvious comment about forecasting with models
using economic variables is that to understand and use such models
properly a knowledge of economics is necessary. At the very least, the
advice of a capable economist should be obtained, for there are many
pitfalls for the unwary in economic forecasting of any sort.
It is useful in considering models involving economic variables,
econometric models, to classify the variables. The usual terminology
defines:
89
(a) endogenous variables — variables occurring in the equations of the

model whose behaviour is ‘explained’ by the model itself;
(b) exogenous variables — variables external to the situation that effect
it but are not affected by it;
(c) predetermined variables — mathematical terms such as trends;
(d) lagged variables — past values of variables (a) and (b).
By way of example, suppose we try to fit a model that relates our
firm’s sales to our prices, one or two indexes of income of our buyers
and their economic state and possibly include a trend term in the
model. We could also relate our prices to past demand, etc. Of these
variables our sales and prices would be endogenous variables and the
various indexes would be exogenous variables, as they would not in
general be affected by anything in our situation. The trend would be a
predetermined variable.
A problem of most econometric models is that we usually wish to
forecast the endogenous variables, but yet the two types of variables
tend to be mixed up together in rather involved regression models. For
example, if A and B are exogenous and x and y endogenous, we might
have
yt =
P0 +
Pi x
t- 1 +
02 3T- i + €t
xt = yt + At + Bt
Such equations do not lend themselves to simple direct solution by
least squares. Much of the literature of econometric models is
concerned with the least squares solution of models of such involved
types. We will not study this topic further here other than to give, in
Figure 6.2, a visual presentation of such equations that is of practical
use. Assuming that the parameters are estimated, we still have the
quantities At and Bt to deal with. In some situations we will be asking
Figure 6.2 An arrow scheme

90
questions about what xt+ i will be if At+\ and Bt+i take on various
different values. These would be referred to as ‘conditional’ forecasts.
If, however, we have to use some forecasts of At+1 and Bt+X to obtain
our forecast of xt+ i, we would have an ‘unconditional’ forecast. A
practical aspect of most econometric models is the large number of
equations involved. The paper by Ball and Burns (1968) gives, an
example of such a set of equations for a relatively small model, together
with a discussion of the forecasting aspects of such models.
A particular problem occurs in economic forecasting, and in general
in forecasting with regression models, that the forecaster needs to be
aware of. The problem is usually referred to as the problem of
multicolinearity. In essence the problem is this: if in the model
+ + e
yt =$0 $\X\t t t
it so happens that x2t v> for all U 7 being a constant, then the
normal equations will not give a solution. The reason is clear if we
substitute to get
+
yt =Po VX11 +et
where 77 = fi1 + y/32 • The equation for yt really only has two parameters,
but yet we were originally trying to solve it for three. In practice we
will not get such an exact relation between the two variables, but we do
often get very high correlations. Though in these cases we can solve the
normal equations, the variances of the estimators we get can become
unacceptably high. One way out of this dilemma is to try and replace
xx and x2 by a new single variable that combines their effects. For
example, in a regression relating sales to temperature and sunshine a
variable termed ‘weather’ could be introduced which combined both
factors in an acceptable fashion. It is never very clear in practice just
how large will be the errors caused by multicolinearity effects. The
safest approach is to use the fitted regression model on new data and
examine the forecast errors.
6.5 Time as a regressor variable

The general changes in variables showing as trends or seasonal variations
have sometimes been regarded as a great nuisance in regression analysis
and only data adjusted to remove these effects have been used in
regression. However, the inclusion of trends and harmonic terms such as
illustrated in forecasting situation (b) of section 6.3 is a perfectly
acceptable device. If the other regressor variables do not show any such
trend or seasonal, the interpretation of the terms will be quite clear. A
problem does arise, however, when a regressor variable also follows its
own trend or seasonal pattern. Suppose, for example, that it follows a
trend with a number of clear divergences from the trend which
91
influence the dependent variable. When the model is fitted by least

squares or discounted least squares, the coefficient of this regressor
variable will adjust to give the best explanation of the effects of these
divergences on the dependent variable. However, this coefficient will
also ensure that the trend in the regressor variable will contribute to the
trend in the dependent variable. The term on the right-hand side put in
to allow for trend will thus only need to account for the additional
trend required to get the appropriate magnitude of trend for the
dependent variable. For example, consider the model
yt = a + fit + yxt + et
Suppose xt has a trend of about 1 unit/unit of t, together with
deviations that are reduced by a factor of V2 in their influence on
Suppose that yt has a trend of about 2 units/unit of t. In fitting this
model 7 will have a value of about V2 to get the effect of the deviations
right. This will thus contribute a trend of V2, so @ will have to be about
% to get the slope of 2 correct. Thus though we will get a good fit, we
will not get a fitted model with a clearly interp re table trend term.
An alternative is to use not xt but a detrended series in which we
have fitted a trend to xt and then subtracted it, leaving only the
deviations from trend. The inclusion of xt means that the estimates of
the trend components may not be so good as if the trend were fitted on
its own. This suggests that, if the time function is a dominant feature of
the data and the other regressor variables, economic variables, say, do
not show this feature strongly, then a reasonable approach is to fit the
time function first. The residuals of this, which in the above are the
‘detrended’ series, are then regressed on the other regressor variables.
Thus, for example, in the above we would first fit
yt = a + pt + r\t
and obtain the residuals
These would then be fitted to xt, or to a similarly detrended xty by

f\t = yxt + et
giving 7. The final fitted model would be
yt = oc + fit + yxt
This fitted model does not have the same estimated parameters as
would have been obtained from applying least squares to the original
model. This method, called stagewise least squares, is discussed by
Draper and Smith (1966) and, in a forecasting context, by Spencer,
Clarke and Haguet (1961).
Following the remarks made in section 6.3(e) on the possible use of
92
\
more involved functions, it is perhaps worth noting that many processes
do show quite involved variation in time. It is well worthwhile
considering which are the best ways of modelling such behaviour. The
method of characteristic modes discussed in Farmer (1964) and
Ivakhnenko and Lapa (1967) provides an approach to choosing and
using appropriate functions of time for constructing forecasting models.
A treatment of this theory is beyond the scope of this book, but it
should be noted that it emphasizes the value of using linear models, so
that we have a model of the general form:
y, = 1 (t) + 002 (f) + 6j
where 0/(0, * “ 1, 2, . . . , are appropriate functions forming the natural

‘modes’ of change of the variable. It also brings out the value of
choosing functions for 0x(f), 02 (0> etc., that are totally distinct in
behaviour (in theoretical terms — orthogonal). There is no point in
having a feature in the shape of 02 (0 which is already in (p1(t).
This brief discussion of time as a regressor variable will be developed
further in the chapter on seasonal forecasting (Chapter 8).
References
Allen, D. M. (1971). Mean square error of prediction as a criteria for selecting
variables. Technometrics, 13,469—475.
Ball, R. J. and Burns, T. (1968). An econometric approach to short run analysis of
UK economy 1965—66. Op. Res. Quarterly, 19, 225—256.
Beale, E. M. L. (1970). Note on procedures of variable selection in multiple
regression. Technometrics, 12, 909—914.
Draper, N. R. and Smith, H. (1966). Applied Regression Analysis. John Wiley and
Sons. New York.
Farmer, E. D. (1964). A method of prediction for non-stationary processes.
Proceedings of Second Congress of the Intern. Fed. Auto Control. Butterworths.
London.
Ivakhnenko, A. G. and Lapa, V. G. (1967). Cybernetics and Forecasting
Techniques. Elsevier. New York.
Robinson, C. (1971). Business Forecasting. Thomas Nelson and Sons, London.
Singleton, P. (1971). Multiple Regression Methods in Forecasting. Thesis M.Phil.,
London University.
Spencer, M. H., Clark, C. and Hoguet, P. W. (1961). Business and Economic
Forecasting in Econometric Approach. Irwin, Homewood, Illinois.
Theil, H. (1965). Applied Economic Forecasting. North-Holland Publishing Co.,
Amsterdam.
Woy, J. B. (1965). Business Trends and Forecasting — Information Sources. Gale
Research Co., Detroit.
Chapter 7
Stochastic models
7.1 Introduction
We consider in this chapter the use of stochastic models in forecasting,
that is to say, the use of models in which the random element plays the
dominant part in determining the structure of the model. In previous
models the chance element et was simply an added ‘error’ added
separately at each time moment to a strictly deterministic function.
Appendix A brings together some of the terminology and definitions of
stochastic theory as relevant to forecasting. The reader should,
however, be able to follow most of the results of this chapter at an
intuitive level.
Consider now the model
Xt = (j)Xt _ i + €t
where the independent chance elements et are identically distributed

with zero mean and variance a2, and 0 is a parameter. This type of
process is called an autoregressive process, since its form represents a
regression of xt on xt_x. To illustrate the structure of the model,
Table 7.1 shows the simulation of data from this model.
If we rewrite the expression for one time unit back,
e
Xt_ j = (f)Xt_ 2 + t- 1
and substitute in the first expression, we get

X, = e, + <pet_ , + 4>2X,_ 2
If we repeat this process over and over again, we eventually get
x, =e, + 0ef_ ! + . . . + <pr~ 2 e2 +4>t~lel + <t>rx0
where x0 is the initial value of the sequence of x. Thus each xt is a
weighted sum of all past e and the initial x0. If t is large and
— 1 < 0 < 1, we may regard the sum as approximately infinite and
93
94
Table 7.1 Simulated autoregressive process

0 = 0.8, x0 ” 1
t <J>x e
t Xt = + €t
t-1
0 1.0
1 0.8 -0.465 0.335
2 0.268 -2.120 -1.852
3 -1.482 -2.748 -4.230
4 -0.384 0.308 -3.076
5 -2.461 -1.178 -3.639
6 -2.911 0.063 -2.848
7 -2.278 0.377 -1.901
8 -1.521 -1.412 -2.933
9 -2.346 1.322 -1.024
10 -0.820 0.204 -0.616
Table 7.2 A simulated moving average process

0 = 0.5
t e# Oet-i xt = et-det_l
0 1.563
1 1.085 0.7815 0.303
2 -0.345 0.5425 -0.887
3 -0.592 -0.1725 -0.419
4 0.399 -0.2960 0.798
5 0.568 0.1995 0.368
6 -0.377 0.2840 -0.661
7 -0.697 -0.1885 -0.508
8 -2.457 -0.3485 -2.108
9 0.410 -1.2285 1.638
10 -0.225 0.205 -0.430
write
xt = et+ 4>e,_ , + </>2ef_2 + . . .
As another example of a stochastic process consider the model
x, = et ~9et_ !
The sequence of x values in this model can be imagined as being a
weighted moving average (or rather moving sum), with weights (1, — 6),
of a sequence of e. This in fact is an example of what is called a moving
average process. Table 7.2 shows an example of this type of process,
again illustrated by showing a simulation.
It is seen that both autoregressive and moving average models can be
expressed in terms of a weighted moving sum (either infinite or finite)
of the sequence of e.
95
It may help if we imagine these models to be formed by a physical

process in which the et values form the input to some mechanism or
system and the xt values the output. Suppose, for example, that when
in a particular system there is only a single unit input at time 1 and no
further inputs, then the system responds by giving outputs of 1 at time
1, 0 at time 2, 02 at time 3, and so on, as in Table 7.3(a). The sequence
1, 0, 02, . . . is often referred to as the impulse response of the system.
Consider the response of the system at time 3 to a random sequence el,
e2, e3. At time 3 the output will be composed of an output of e3 due
to e3 arriving at time 3, 0e2 due to e2 arriving 2 and (p2ex due to ex
arriving at time 1, assuming here that the magnitude of the output is
simply the impulse response times the size of the input. Assuming
further that the actual ouput is the sum of the contributions of the
responses to ail past and present inputs, we arrive at a total output of
x3 = e3 +<pe2 +<p2e1
A system for which we can carry out the above multiplications and
additions is called a linear system and the models so defined are called
linear models. Table 7.3(b) shows how the output sequence for this
particular system is obtained from the input sequence for t = 1, 2,
3, . . . . It will be seen that for t tending to infinity we obtain for the
output sequence the simple autoregressive model.
We have described a physical system in this example. There are many
situations in business and economics where similar phenomena may be
expected. The et quantity is the total effect of the unpredictable
factors in the situation at time t. The quantity et will not only affect xt
but will also influence xt+ i, xt + 2, etc., though the influence will reduce
as time passes, a reduction in our example governed by the parameter 0.
Consider now the problem in our example of forecasting x4 and x5
from the preceding data. The output at times 4 and 5 will depend on
two types of input. Firstly, it will depend on the inputs at times 1, 2
and 3, whose effects we have already observed and, secondly, on the
future inputs e4 and e5. We cannot attempt to forecast any particular
numerical values for these inputs since they are independent of any
existing information. The best we can do is to use their expected values,
which as before we will take as zero. Thus we have to replace the
known structure of Table 7.3(b) with a similar table, Table 7.3(c),
which allows for our ignorance of future events. Thus our forecasts of
x4 and x5 will be
x4 = 0 + 0e3 + 02e2 + 03e1
x5 = 0 + 02 e3 + 03 e2 + 04
In terms of the observations xx, x2, x3 these will be
—
X4 0X3
96
\
Table 7.3 Inputs and outputs of a simple linear system
(a) Impulse response
Time 1 2 3 4 5 6
Input 1 0 0 0 0 0
Output 1 0 02 03 04 0s (—1 < 0 < i)
(b) Response to a sequence

Table elements give the response at time T due to input at time t.
Input times t 1 2 3 4 5
Input ei e
2
6
3 e4 e5 Total output at time T
X y1
Output times T
1 €1
2 0e 1 ^2 e2 + 06!
3 026I e 63 + 0e2 + 026I
0e2 3
3
4 0 ei 02e2 0^3 €4 e4 + 0e3 + 02e2 + 036I
5 04ex 03e2 02e3 0e4 e5 e5 + 0e4 + 02e3 + 0362 + 04ex
(c) Response table at time t = 3
Input times t 1 2 3 4 5
Input ei ^2 €3 Total output at time T
Output times T Actual

1 past ei Cl
2 0e 1 ^2 e2 + 06!
2
3 present 0 e1 0e2 e3 e3 + 0e2 + 02 63
Forecasts Forecasts
4 future 0 3
e! 02e2 0^3 0 0e3 + 0262 + 03 61
4
5 0 î 0 e2 02e3 0 0 02e3 + 03e2 + 046I
and
Xs = <t>2X3
Now we have a simple intuitive method of forecasting. Whether or not

this is a good method is not obvious; in fact this gives forecasts with
minimum mean square error for this model. To obtain the forecast
error, we simply note that we have replaced the contributions in the
outputs from e4 and e5 by zero and hence the errors are
e4 ~ e4
97
and
e5 = e5 + 0e4
as will be seen by comparing parts (b) and (c) of Table 7.3. The error
variances, which here equal the mean square error as the expectations
are zero, are thus
E(el) = o2
£(e2) = (l+02)a2
again using the independence of the e values.
It should be clear from the above special case that for the
autoregressive model the forecast with lead time h is
xt+h =<j>het + 4>h+iet_ I + 4>h + 2et_2 + • • •
which is
xt+h = <t>hx,
We may also derive this formula intuitively from the basic form of the
model. The future observation xt+ 1 can be written as
x —
t+ l £f+ l
At time t, xt is known but et+ i can at best be forecast as zero, so we

have as a forecast
—
%t + I
Similarly, the future value xt+2 can be written as

x =< x +
t+2 P t+l ^t+ 2
At time f, xt+ i is not known but can be forecast by xt+ x; et+ 2 again is
forecast as zero. Thus
=
%t+ 2 0*^f + 1
and so
Xf + 2 0
Thus, in general,
Xf+h 0 Xf
For this forecast the error et+h is given by
+ h ~ Cf + h 0C?+ h — 1 ^"•••*^’0
since the error consists of the part of the model of xt+h involving the
future unpredictable e. The error has zero expectation so that the
98
\
forecast is unbiased. The error variance is given by
Var(ef + h ) = o2 (1 + 02 + . . . + (p2h ~ 2
)
and this can be used to calculate the necessary confidence limits for the
forecast. If o2 is not known, we must consider how to estimate it. This
can be done by noting that if we forecast only one step ahead, h = 1,
then
e e
t+ I t+ I
If we go back through our past data and use the forecasting formula
Xt+ 1 (])Xf
then the set of forecast errors obtained will in fact be the actual
random variable, e. The variance of the e value is thus o2 and an
estimate of a2 is provided by
„
2
1 t
d =- X ej
t t= I
Table 7.4 shows the forecasting procedure for the autoregressive

model discussed above. It is seen from this example that to start the
process we need an initial value x0. In the example of Table 7.4 we
have forged data for which x0 = 1 and 0 = 0.8, but in carrying out the
forecasts we have assumed 0 to be known to the forecaster but x0
unknown and set at the value zero, just for the sake of illustration (in
practice neither 0 nor x0 will be known). In the example the effect of
the starting value dies away immediately and the forecast errors are
equal to the e values of the simulation. The real problem then is not of
choosing x0 but of estimating 0. The method of least squares can be
Table 7.4 Forecasting an autoregression process

Known model parameter 0 = 0.8
True initial condition x0 =1.0
Initial condition for forecasts x0 = 0
Simulation of data Forecast Error

(pxt- I e
t xt X -f-jf- j (px j-
1.0
0.8 -0.465 0.335 0.0 0.335
0.268 -2.120 -1.852 0.268 -2.120
-1.482 -2.748 -4.230 -1.482 -2.748
-3.384 0.308 -3.076 -3.384 0.308
-2.461 -1.178 -3.639 -2.461 -1.178
-2.911 0.063 -2.848 -2.911 0.063
-2.278 ' 0.377 -1.901 -2.278 0.377
-1.521 -1.412 -2.933 -1.521 -1.412
99
used for this purpose. The value of 0 is thus chosen to minimize the
sum of residuals squared, i.e.
2 ef = 2 (Xj — ! )2
i= 2 i—2
The minimum is obtained when
0 = 2 XjXi_ ! 2 xf_ j
/= 2 i=2
The summation starts from i = 2 rather than i = 1, as x0 is not an

observation. As new data are obtained, both numerator and denomin-
ator can be modified by addition of the next available product and
squared term (xt+lxt and xf), respectively. If we wish to regard our
basic model as a local model, we may estimate 0 using the method of
discounted least squares. This has the effect of replacing the sums in the
formula by weighted sums. The discounted estimate with exponential
weights can be written as
/t — 2
r
4>t= 2 a xt_rxt_r_l / 2 arx2t„r„ , = Nt/Dt
r= 0 y= 0
We may update both numerator and denominator by the usual type of

recurrence relations:
t- I t-2
2 CL X f + i _ r Xf +1 _ y — i X t+ 1 x, + a 2 a'A, x f- r- i
r-0 r=0
i.e.
Nt + i = xf+ i xt + aNt (Nx = 0)

and
t-1 t-2
2 r
a x?+1_r_! = x? +a 2 arx]_r_x
r=0 r= 0
SO
Df+1 = x] +aDt {D1 = 0)
Table 7.5 illustrates the use of this method on the same data as that
in Table 7.4. It will be seen that the initial estimates of 0f are wild, but
soon settle down to vary about the true value used in the simulation.
Similarly, the forecast errors soon become fairly close to the e values
used in the simulations of Table 7.1. We have taken the discounting
factor a as 0.8 to illustrate the method, but clearly we are dealing with
simulated data which we therefore know to come from a globally valid
100
\
Table 7.5 Forecasting the data of Table 7.4
0 estimated using discounted least squares, a = 0.8
Forecast Error
xt *t- 1 N, D, 0* (pt%t e
t
0.335
-1.852 0.335 -0.620 0.112
-4.230 -1.852 7.338 3.520 — — —
-3.076 -4.230 18.882 20.709 0.91 — —
-3.639 -3.076 26.299 26.029 1.01 -2.799 -0.840

-2.848 -3.639 31.403 34.065 0.92 -3.675 0.827
-1.901 -2.848 30.536 35.636 0.86 -2.620 0.719
-2.933 -1.901 30.005 30.866 0.97 -1.635 -1.298
-1.024 -2.933 27.007 33.295 0.81 -2.845 1.821
-0.616 -1.024 22.237 27.685 0.80 -0.829 0.213
model. Thus a = 1 would be the correct value to use; it would give the
best estimator of 0 and the best forecasts.
In this section we have concentrated on developing some intuitive
ideas about forecasting stochastic processes. To progress further, we
need a method for justifying these intuitive methods.
7.2 Forecasting and conditional expectation

In this section we discuss briefly how to find a forecasting formula,
given observations xx, x2, . . . , xt, and the stochastic, or other, model
underlying the data. Clearly, we would wish the forecasts obtained from
such a formula to satisfy some criteria of quality. The one we will use is
the mean square error. Thus we require an expression for the forecast
xth ofxt+h for which
MSE = E{(xt+h — xtfh )2 }
is as small as possible. The basic result that we will use is that the xt>h
which minimizes the mean square error is the expected value of the
future xt+h, given that x{, x2,...,xf are already known. This is
referred to as the ‘conditional expectation’ and is denoted by
E(x f+ fa I Xi > X2 , • • • 5 )
This result is derived in Appendix B. To see how to use this result, let
us look again at the autoregressive model of the last section, namely
xt+ i = $xt + et+1{t = 1, 0, ... )

We have written the model for xt+ x since the result above states that
the forecast of xt+ x is its expectation, given xl9 . . . , xt. Formally
101
Xft i E(Xf+ j I Xj , . . . , Xf)
= E(cpxt + et+1 \xl9 . . . , xf)

The conditional expectation of xu given the value of xt, must be that
given value, xt. As the future e values are independent of past events,
the conditional expectation of et+ x, given xl9...,xt9 is just its
ordinary expectation, which is zero. Thus
—
Xft i (j)Xf
which was the result argued intuitively in the last section.
7.3 Moving average processes

In section 7.1 we mentioned the model
xt = et — det_ 2
This process is an example of the moving average process. The general

expression for such a process is
Xf ~ € t 0i _ i 6 2^ t— 2 • • • @ q^t — q
where 6 j, 02 , • • • > 0q are constants. This will be referred to as a finite
moving average process of order q, MA(q). The series might continue to
infinity, in which case it would be termed as an infinite moving average
process. We assume that either the expectation of x is zero naturally or
if it was originally p we subtract p from each actual observation to
obtain x. The forecasting of processes like this and other related
processes to be discussed later in this chapter has been investigated in
detail by G.E.P. Box and G.M. Jenkins (see Anderson, 1975; Box and
Jenkins, 1970; Nelson, 1973); on the whole the notation of this chapter
is that used by Box and Jenkins.
Let us start by considering a simple moving average process of order
one:
xt=et~ 6et_ i
The minimum mean square error forecast of xt+ x is
Xf ~ E(Xf+ 2 ^ Xfy Xf _ 2 5 • • • )
As fixing the values of xt9 x?_ i, . . . is equivalent to fixing the values of

et, et_ 2, • • • , we can rewrite this as
Xf ~ E(xt+ 2 I 6fy €f — 2 5 • • • )
—
E(€f+ 2 I t
^ — 1 5 • • •)
= E(et+1) — det
102
>
The future et+l has zero expectation and E(et \ et) is the expectation
of a constant, et. Thus xt = — 0eti and we obtain a formula for the
forecast in terms of et. To obtain a more practical forecast, we first
examine the prediction error
—
@t+ 1 •%t+ 1 ^t
= (e,+1 -0ef)-( —0et)

=
£f + 1
We again have a situation in which the one-step-ahead forecast errors,

which we can obviously determine directly, are identical to the e values
in the model. The forecasting formula can thus be rewritten as
xt = — det
which is also expressible in the recurrence form
xt = — 0(xt —xt_ i)
If we seek to express xt directly in terms of the data, we may
repeatedly substitute for xt_i, xf_2, ... in the last expression,
giving
xt = — 6xt — 62 xt_ i — 63 xt_ 2 — • • •
It follows from this last expression that to make sense of our forecast
we must require that 6 lies in the interval (—1, 1). Finally, as et = et the
minimum mean square error is in fact o2. Table 7.6 shows a set of data
and forecasts for this model, in which e0 has been taken as zero to
provide a starting value.
Suppose now that we wish to forecast two time units into the future.
Table 7.6 Forecasting a moving average

process
5ct = ~det, d = 0.5, e0 = 0
t xt %t-l e
t xt
0
1 0.303 0 0.303 -0.1515
2 -0.887 -0.151 -0.736 0.368
3 -0.419 0.368 -0.787 0.393
4 0.798 0.393 0.405 -0.202
5 0.368 -0.202 0.570 -0.285
6 -0.661 -0.285 -0.376 0.188
7 -0.508 0.188 -0.696 0.348
8 -2.108 0.348 -2.456 1.228
9 1.638 1.228 0.410 -0.205
10 -0.430 -0.205 -0.225 0.112
103
From our model

x
t+ 2 = e
t +2 @et+l
on the right-hand side of which we have only future values of e. It

follows that
E(xt+ 2 Iff, ef_i. . . ) =E(et+2) — 6E(et+i) = 0
Thus we are unable to obtain a better predictor of xt + 2 than its mean
value of zero, and the same obviously applies to any JC more than one
time unit into the future.
The generalization of the above example to obtain forecasting
formulae for the moving average process of order q is quite straight-
forward. The forecast is given by the expectation of xt+x conditional
on xt, xt~ i, . . . , which is equivalent to conditioning on ety et _ x, . . .
Hence
Xt~E(€f+i ^ l 2
@ ^t — 1 • • • @ q ^ t+ 1 — q ^ ^t— 1 > • • • )
SO
Xt ~ 0 i €t d 2^ t — 1 • • • @ q ^ t+ 1 — q

x x
@t+ I ~ t+ i t
which, on expressing both xt+ { and xt in terms of e values, gives

e =
t+ I £f+1
So once again the one-step-ahead forecast errors correspond to the e

values and the minimum mean square error is o2. Using this fact to
obtain a more practical form of forecast, we have
Xf ~ 2
^ ^t—\ • • • @ q ^t+ l — q
where the e values are past forecast errors.

The standard deviation of the prediction error is o and so confidence
limits for xt+iy given our forecast, can be obtained directly. For
example, the 95 per cent confidence limits would be xt - 1.96 a,
assuming a normal distribution. Where a is not known, it is reasonable
to estimate it directly using the sample standard deviation of the
forecast errors.
If we now seek to predict two steps ahead, the argument used before
gives
X -f 2 0 2 ^t 03 £f l ... 0q Ef + 2 q
or
Xf 2 02 ^ 3 &t 1 • • • ^ q ^t+ 2—q
104
where the e values are still the one-step-ahead forecast errors. Compar-
ing the forms of xt and xu2, it is seen that if we are forecasting both
xt+ i and xt+2 at each stage we can simply update our forecast xt using
x + x
t+ 1,1 ~®\et+\ t, 2
We can continue generating forecasts until we seek to forecast

xt+q+ i for which the basic model gives
x
t+q + 1 ~ ^ t + q + 1 ^ 1 ^t+q ••• 0q €f+ 1
As there are only future e values on the right-hand side, the conditional
expectation of xt+q+ x is simply zero. Thus we have
x
t,q + k=0 (k>l)
The forecast error for forecasting h{h < q) steps into the future is
x x
@t+h ~ t+h t, h
~ i.^t + h ^ 1 ^ t+ k + l ... @ h — l^f+1 ^ h^t ••• Oq^t+h — q)
( @ h^t ••• @ q ^ t+ h — q)
~~ ^t + h @ 1 ^t+ h — 1 ... @ h — 1 ^t+ 1
i.e. it equals the ‘future’ part of the model for xt+h. This has zero
expectation and variance
°uh = o2 (1 + 6\ + ... + Oh - 1 )
If we wish to give ‘confidence’ limits for xt+h, given our forecast, we

have, for example, that the 95 per cent confidence interval, or
prediction interval, is
x
t,h ~~ 1*96 otth
assuming normality in the e values.

The above discussion was concerned with models only. To make this
method practical, we need to be able to estimate the parameters
6!,..., 0q. The theory of this estimation is beyond the scope of this
book, but is dealt with in detail in the book by Box and Jenkins, using
the method of least squares. The use of discounted least squares is also
possible. The basic problem with both methods is that the model does
not enable the residual sum of squares, Se2 (discounted or otherwise),
to be written as a simple function of the parameters. Thus we have
nothing corresponding to the normal equations that can be simply
solved. The methods needed are thus fairly complicated and involve
repeated approximations to the answer. The methods do not enable the
parameters to be re-estimated, as new data is obtained, in a simple
recursive fashion.
105
By repeated substitution using

e
t+1 = e
t+ 1
= xt+ I xt
in both models and forecasts for moving average processes, we find that
both can be expressed as an infinite series of past observations, whose
coefficients depend on 6 j, . . . , 6q. Experience suggests that with many
sets of data the first few terms of these series can be used to provide
good forecasts. We can thus use an autoregressive form of model as an
approximation to a moving average model. This avoids the problems of
estimation referred to above, since, as we have already illustrated in
section 7.1, the autoregressive parameters can be directly estimated.
7.4 Autoregressive processes

The general pth order autoregressive process is defined by
Xf 01 Xf_ j 02*^?— 2 * • • 0p*^'f — p ~
where 0j > 02, . . . , 0P are constants and the model is denoted AR(p).
To investigate the forecasting of such a process we start with the
simplest case, the first-order autoregressive process, which we have
already studied in some detail in section 7.1. The model is
=
*t 0\Xt- I + et
We have already shown in section 7.2 that the minimum mean square
error forecast of xt+ x is
Xf 0 j Xf
^t+ 1 ~~ Xf+ i xt
= {4)1xt + e(+1) — <t>ix,
= et+1
So the forecast error is the future et+1, and the mean square error is
thus a2.
To forecast two steps into the future, we obtained in section 7.1
=
Xf ,2 0i xt
As a somewhat different way of getting this result we could write
Xt ,2 ~ -^(01 *^<'+1 ~*"^f+2 \ X f y Xf _ !,*••)

— 0 1 E(xt+ i |x^,Xf_j,...)
which is
=
Xt, 2 01 Xf
106
\
and hence
x x
t,2 ~ 01 t
The prediction interval may be found directly by writing

er+2 = (<p}xt +</>,e(+1 + ef+2) — <j>\xt
=
01 e
t+ 1 + €t+ 2
So
Var(ef+2 ) = ofA = (<t>] + 1 )o2
Hence, assuming normality, the 95 per cent prediction interval for xt+2
is
<j>\xt - 1.96(0i + l)a2
For the general pth-order autoregressive processes we have
x
t+l = <t>\Xt + <S*2Xt- 1 + • . • + 0pxf+ i _p + et+ i
and so, as
—
xf i E(Xf+ i | x^, . . . , Xj)
we have
= 0i X* + <t>2Xt- 1 + • • • + $pxt+ 1 -p
as the one-step-ahead forecast and

6
£*+I ~ #*+ I ^/,i ~ r+ I
as the one-step-ahead forecast error. Thus the model provides the

forecast in a very direct fashion and once again the e values are equal to
the one-step-ahead prediction errors.
For forecasting two steps ahead,
—
^ t, 2 E(Xf+2 I x
ti • • • > x
\ )
x X x X
^(01 t+ 1 $2 t 03 t— 1 0p + 2 — p + ^t+ 2 \ ti ’ * • > *^T )
— 01 E(X{ + i I X^, ...,Xi)-^02X^"^..."^'0pX^4.2 — p
X x
t ,2 ~ 01 1 02 t &pxt + 2 — p
By a similar argument
x
t,3 ~ $lxt,2 &2xt,l 03 X
t *ftpxt+ 3 — p
Thus in predicting a future value xt+h we first write down the model
for xt+h and delete the term e . When an x in the model refers to a t+h
future value, we replace it by its predicted value, x; where it refers to a

past value we simply use the observed value. To find the prediction
107
errors of the above, we subtract xu2 from xt+2 , giving

e = e + (
t+ 2 t+ 2 P\(xt+l x
t,l) = e
t+2 +
4>let,l
and subtract xt 3 from xt+ 3, giving
, 3 ~ €f+ 3 "*"01 (*Vf + 2 x

t,2 ) + $2 (x
t+l x
t, 1 )
= e +( ) e
t+3 ti t,2 + (p2eti
and so
=e + + +
f+3 01^t+2 01^f+l 02et+l
We can use relationships such as this to find prediction intervals for

different lead times. Thus
Of,2 = o2 + Var(et>1) = 2
CT (1 + 0? )
and
Of,3 = 02{1 + 4>\ + (01 + 02 )2}
The above discussion assumes that we know the parameters <j>x . . . ,

0p. To obtain practical forecasts, we must clearly obtain estimates of
these from the data. As the model is of linear regression form, we may
use the methods of least squares or discounted least squares described
in the previous chapter to obtain appropriate normal equations. A
simple illustration of this was given for the model AR(1) in section 7.1
and illustrated in Table 7.5.
7.5 Autoregressive — moving average models
Box and Jenkins (1970) have considered a generalization of the two

previous models. The reader is referred to this reference for details. The
aim here is to illustrate the basic ideas of this more general model by
considering some of the simpler cases.
The most natural generalization of the two previous models is to
combine them. Thus if we take the first-order moving average process
x
t ~ ^t @ — I
and the first-order autoregressive process

x
t &xt — I ~
and combine their structures, we obtain

x
t 0*Vf— 1 ~ ^t 0€f_ i
This would be referred to as a ‘mixed moving average — autoregressive

process’ of order (1, 1).
108
To forecast xt+1 we note that

—
Xf+ i ^t+ 1 0€f
So taking the conditional expectation at time t gives
xt = (j)xt — 0et
£t+ i *^f+1 ^t ^t+ I
so once again the one-step-ahead forecast errors are equal to the

corresponding e values. Thus for practical use we may write
xt = (pxt — 6et
or in terms of the previous forecast
Xt = (0 — 6)xt + 0xt_ !
If 0 = 1 and 0 < 6 <1, this is identical to the forecasting formula of
exponential smoothing; thus exponential smoothing is the minimum
mean square error forecasting method for the model
Xf Xf _ \ 0€f__ i
Here each value of xt equals the previous value xt_ j plus a random
term from a simple moving average stochastic process.
Consider now a prediction for a lead time of two in the original
model
—
% t ,2 E((j)Xt + 1 2 @ 1 \ %t9 • • • >)
= (f)E{xt+ ! I xt . . . )
—
0**-1,1
This same calculation obviously applies to any lead time, k say. Thus
xt,k =<l>xt,k- 1 (k>2)

so
Xt,k = <Pk~ 'xt,l
Considering now the problem of updating the forecasts, substitute for

xt i in this last expression; then
xt,k =k-l(<l>-e)xt + 6<t>k-lxt-ltl

so
k 1
xt,k = 9xt- 1 > + 4> ~ {<t> — 0)x,
Thus there are very simple recurrence relations for updating forecasts.
109
As with autoregressive and moving average models, we may genera-

lize the above model to form
x
Xt 01 t— 1 02 — 2 • • • 0p %t— p ~ £ t @ \^t — 1 @ 2^t - 2
— ... — f)v fc _
q f q
This is referred to as an autoregressive moving average model of order

(p,q), denoted by ARMA(p,q). The general procedure for forecasting
with such a model follows exactly the lines indicated above.
At first sight a mixed model may appear to be a rather complicated
statistician’s toy without much intuitive basis. It can be shown,
however, that we can expect such models to be appropriate in a range
of practical situations. These include
(a) situations where the series of interest are an aggregate of a number
of other series following simpler stochastic models;
(b) situations where a simpler model is observed subject to error;
(c) situations where an original simpler series is aggregated over time,
e.g. where weekly data are aggregated to give monthly or yearly
totals;
(d) situations where the series of interest are subject to feedback
effects through their interactions with another variable.
Let now summarize the rules for writing down forecasting formulae
for moving average, autoregressive and mixed models. To forecast xt+ k,
write out the expression for xt+k as given by the model.
This will be a linear function like
Xf + ic terms in Xf+k — ^, . . . ^Xf+i^Xf^Xf — j, . . . , €. ^+ k, . . . , e^+j,
— I> • • •
then xt> k is obtained directly from the right-hand side of this by

(a) replacing all future e values by 0, i.e. et+ r = . . . = et+ k = 0;
(b) replacing all past e values by the corresponding forecast error of
lead one, i.e. eT by eT = xT — xT_ i i,
(c) replacing all future x values by their forecasts, i.e xt+h by xt h;
(d) leaving all past x values at their observed values.
Example Consider prediction for the model

xt — 0.5xt__i +0.3X?_2 =6f-0.8
(a) Thus
xf+1 = 0.5 if — 0.3 xf_i + ef+i 0.8 et
so
octti = 0.5 xt — 0.3 xt_x —0.8et
110
\
x =
(b) t+2 0.5Xh1 — 0.3 xt + et+2 — 0.8 et+1
so
=
xt2 0.5 xt i — 0.3 xt
(c) Xf+fo ~ 0.5 _j 0.3 xt+h_2 "t* £f+0.8 €f+h_i(k ^ 3)
so
x
t,H ~0.5x^ ^_i 0,3 Xj/j_2
Using the extreme simplicity of these relations, we may also derive
expressions for the forecast error variance for different lead times. Thus
subtracting the forecast formulae from the models gives
(a) et+1 - et+1, for a lead time of one,
so
= a2
(b) For forecasts xt>2, with lead two, the error is

=
&t + 2 ^t, 1 ^t + 2 €f+ 1
= —0.3 et+1 + et
so
o] 2 = 1.09 a2
(c) For a lead time of k
€f+h ~ ^t+h — 1 0.3 @t + h—2 £t + h 0.8 ^-t+h —X
Here et+h_1 is not independent of et+h_2 and so we cannot
simply relate a2>h to a2>k_ x and a?By repeated use we may
relate cf+ 3, et+4, etc., to the e values and hence find a2 3, a2?4, etc.
As usual we have assumed known parameter values. These models

pose difficult problems of estimation, though computer programs to
carry out the estimation are becoming increasingly common. Unfortu-
nately, few of these programs update the estimates of the parameters
as new data become available. They tend to keep to the parameters
estimated from the original set of data, but include some form of
quality control of the forecasts to indicate when the model may have
changed. Experience, and intuition, suggest that updating the parameter
estimates as new data arise lead to better forecasts. It should also be
noted that where discounted least squares is used forecasting para-
meters of 0.95 and above commonly produce the best forecasts. We
have also assumed in the discussions of this chapter that we have
identified an appropriate model for the data in hand. The identification
Ill
of stochastic models requires considerable data and experience and the

details are beyond the scope of this book. Section 12.2 and appendix A
give some indication of the approach but the reader is referred to the
books by Box and Jenkins (1970) and Anderson (1975).
7.6 Models involving stochastic and deterministic elements

In the previous sections we have been examining stochastic models
which contained no deterministic part. This is not very realistic, as at
the least the data will have some non-zero mean which may or may not
be globally constant. We therefore need to introduce methods for
dealing with situations that show both stochastic structure and
deterministic elements such as means and trends. There are two main
approaches that have been introduced for such situations:
7.6.1 methods based on the elimination of deterministic elements
from the data before the use of stochastic models;
7.6.2 methods based on the formulation of models that, though
stochastic and global in nature, correspond to local deterministic
structures.
Let us illustrate these two approaches in turn.
7.6.1 Elimination of deterministic elements

7.6.1 Suppose we look at a simple linear trend model
xt = a + (3t + et
The trend part of this model can be eliminated by a process of
differencing. That is to say, by using the differences between successive
values, an operation that we denote by V. Thus differencing the data
once gives
Vxt = Xt — Xt_ !
= /3 + et — et _ i, in terms of the model
Repeating this, and using V2 xt to denote the repetitions, we get

V2 xt = Vxt — Vxf_!
= et — 2 et_ i + et_ 2 ? from the model
Thus if we define a new variable wt by

wt = V2
the model becomes
wt = et — 2 ef_ ! + €t-2
112
The model for wt is therefore a moving average model of order two.

Thus the structure of a liriear trend model can be incorporated into a
moving average model by the use of differenced data V2xf in place of
the actual data. This suggests that very general models of a non-
stationary nature can be constructed from moving average models,
autoregressive models and the mixed moving average—autoregressive
model by replacing the xt by differences of xt, say the dth difference
Vdxt. Such models are termed ‘integrated’ models of order (p, d, q),p
referring to the order of the autoregressive part of the model, d to the
degree of differencing and q to the order of the moving average part of
the model.
Consider, by way of further illustration, the integrated moving
average model of order (., 1,1), i.e.
Vx, = e, — 0ef_ i
Writing this as
X, = xt_ 1 + et — det_ j
gives a form identical to the particular model discussed in section 7.5.

As any difference operator Vdxt can be expanded in terms of xt,
xt_ xt_d, it is clear that any integrated model can be expressed
in autoregressive terms.
As the differencing uses only adjacent values of x, it is unaffected by
slow variation in the underlying mean. Similarly, a model using V2xt
will be almost unaffected by slow variation in both mean and trend.
The practical procedure here is to difference the data until a stable
series of zero mean and small variance is obtained. If the series is
differenced too many times, the variance will increase and the process
of identifying and fitting the model for the differenced data will get
very difficult.
The use of differencing may in practice suffer from two disadvan-
tages. Firstly, to obtain the forecast of a future xt+i, for example, we
need to forecast Vif+1 by V xt+ x, and then ‘undifference’ to be able to
write
=
Xf+i Vif+ i + xt
Thus xt contains the measure of underlying mean, since we are not

explicitly estimating pf, the mean at time t. This may increase the root
mean square error above that of a method that explicitly estimates the
local value pt. As always, this can only be tested by trial and error. The
other disadvantage is that often knowledge of the local mean and trend
is an important side product of the forecasting exercise. This knowledge
is not provided when differences are used. This disadvantage is
overcome in the second approach to this problem.
113
7.6.2 Stochastic models with deterministic elements

The second approach seeks to formulate the model in a form that
explicitly includes the deterministic parameters. Three particular
models are of value here:
(i) If E(xt) = jU, for all time, then all the previous stochastic models
can be straightforwardly reformulated with (xt — ju), (x^_ x — /u), etc.,
replacing xt, xt_x .... Estimation of ix is not very difficult, given a
good quantity of data.
(ii) If a constant trend occurs with an autoregressive model, it can be
allowed for by the addition of a constant term in the regression form.
Thus, for example,
xt = 01 xt- i + et
would include a slope /3 if it was modified to
xt = @ + j + e,
Again the estimation and forecasting procedures would be virtually

unchanged by this additional term.
(iii) The third type of model gives a formulation that is stochastic
and global in nature but which allows for local variation in mean and
slope. We start with the ‘local’ constant mean model
xt = Ht + et
where
E(et) = 0 and Var(ef) =
We dealt with this model in section 4.2 by assuming that \xt wandered
in an unknown, undefined, fashion. We now formulate a model that
describes how fxt changes. If there is no trend, we use
= +
Mr Mf- I 71
where
E{ 7r) = 0 and Var(7,) =

This represents a mean that is constantly modified by the random
variations, yt. It can be shown that, for large t, the forecast leading to
the minimum mean square error for this model is just exponential
smoothing. The forecasting parameter a of the exponential smoothing is,
however, no longer an empirically introduced parameter; it is related
directly to the model parameters o\ and Oy by
a = [(1 + 2R)- /(I +AR)]I2R {R = o] /e*)
Thus, for example, if approaches zero, then a approaches 1, which

corresponds to the local model approaching the global case of a fixed p.
114
The values of o\ and a2 can be estimated from the correlation structure

of the original data (see Grunwald, 1965).
The above model can be extended to allow for trend by using the
ideas developed for Holt’s method and illustrated in Figure 5.5. Thus
the mean model is rewritten as
Mt =Mf- 1 +0f +7r

and a wandering slope is modelled by
=
Pt fit- I +
where
E(8t) = 0 and Var(5f) = a§

It can be shown (Harrison, 1967) that Holt’s method provides the
optimum (MSE) forecast for large t, for data from this model. Again
the forecast parameters are related directly to the model parameters,
a2, a2 and a2. In a paper by Harrison and Stevens (1971) methods for
estimating and updating the parameters are described. These allow also
for the error variance and the best forecasting parameters to be updated
as the process develops. As these methods depend on the use of a
formula and approach developed by Rev. T. Bayes, over a century ago,
they are often referred to as Bayesian forecasting methods. The
technical details are beyond the scope of this book, but a further facet
of the approach is discussed in sections 10.3 and 14.4.
It can be shown that the above models correspond to particular

moving average models for Vx and V2x, respectively. The difference
lies in the formulation of the models. For the above formulation p and
/3 are explicitly estimated by the forecaster, which is not the case in the
differenced forms.
References
Anderson, O. D. (1975). Time Series Analysis and Forecasting, the Box—Jenkins
Approach. Butterworths, London.
Box, G. E. P., and Jenkins, G. M. (1970). Time Series Analysis, Forecasting and
Control Holden-Day, Inc., San Francisco.
Grunwald, H. (1965). The correlation theory for stationary stochastic processes
applied to exponential smoothing. Statistica Nierlandica, 19, 129—138.
Harrison, P. J. (1967). Exponential smoothing and short-term sales forecasting.
Manag. Sci., 13, No. 11, 821—842.
Harrison, P. J., and Stevens, C. F. (1971). A Bayesian Approach to short-term
forecasting. Op. Res. Quarterly, 22, No. 4, 341—362.
Nelson, C. R. (1973). Applied Time Series Analysis for Managerial Forecasting.
Holden-Day, Inc., San Francisco.
Chapter 8
Seasonal models
8.1 Introductory concepts
8.1.1 Seasonal structures

In this chapter we study forecasting using seasonal models, i.e.
models which show some form of variation over each year which
repeats, at least approximately, year by year. The year may be 52
weeks, 12 calendar months or 13 four-weekly periods. We may also
apply the methods to variation within a working week which repeats
weekly or any other variation which has a known fixed period. The
basic cause of seasonal variations is usually either the seasonal weather,
as in, say, airline bookings or sales of sun tan lotion or cough mixture,
or the fixed seasonal events such as Christmas, motor shows and sales.
However, the cause may be far removed from the series being studied.
Figure 8.1 shows some examples of various types of periodic
variation. A major problem facing the forecaster of seasonally varying
data is that of choosing the right type of model, since there are many
ways in which seasonal structures occur. For the present it is sufficient
to observe from Figure 8.1 that a seasonal structure may be constant
from year to year or it may show some deviation from one year to the
next. This deviation may be in the underlying mean level of the
variation, as in Figure 8.1(b), or in the positioning of the features of the
variation within the year, as in (c), or in the magnitudes of the
variation, as in (d). In practice, of course, the deviation may be of any
or all of these forms. To make the models realistic we must also assume
that random variations also occur in our data.
In this chapter we will survey some of the most frequently used
methods of analysing and forecasting data with seasonal features.
Rather than look at each of these methods separately, we will seek to
emphasize the common structure that most of these methods have and
show the various techniques in these methods as particular approaches
115
116
Figure 8.1 Types of seasonal structure (a) (b) (c) (d)
to solving certain common problems. The one common element in all

the methods discussed in this chapter is their basic assumption as to
what constitutes a general seasonal model. They differ in the way the
model is expressed, but all see in it at least three of four components.
These are explained briefly below.
(a) The trend component T
This is the long-term upward and downward movement of the series

due to slow changes in factors that influence the mean of the series.
Depending on the nature of the series, these factors might be, for
117
example, growth in the firm, economic change, population change and

change in the quality of materials.
(b) The seasonal component S

This is some measure of the characteristic behaviour of the series
during each season in the period, e.g. each month in the year. The
seasonal may be assumed to repeat itself in each period, though often
we allow for slow changes in its structure.
(c) The cyclical component C

This has something of the character of the seasonal component in
that it is expected to show upward and downward swings, but the term
‘cycle’ is kept to describe movements which are slower and much less
predictable than the seasonal component. The length of the cycle and
its peak magnitude are all uncertain. The length of many cycles average
at about 3 to 4 years, though some are longer than 15 years. The nature
and predictability of cycles has been under detailed discussion for many
years. The reader is referred to the book by Dauten (1961) and Davis
(1941) for a thorough treatment of this component. When considering
short-term forecasting, it is convenient to consider the trend and the
cycle together. The reasoning for this is simply that it is impossible to
pick out useful information on cycles without data from several cycles,
which is usually at least 10 years. Any cyclic feature in data obtained
over shorter periods thus gets interpreted as part of the trend
movement. Conversely, the trend itself can be regarded as part of a
cycle of very long duration. In short-term forecasting the cycle is
usually included in with the trend and in this chapter we will assume
that the term T includes C in the short term.
(d) The irregular component I

This is, in effect, defined as the remainder when T, 5 and C have
been accounted for. It will usually be regarded as being a random
sequence. It may, however, be that some of its peaks and troughs can
be explained in terms of the origin of the data. A trough corresponding
to low sales might be due to some breakdown in delivery and the
following peak due to the making up of the backlog. A valuable part of
the analysis of data using this model is the increase in understanding of
the situation which comes from this type of investigation of possible
cause and effect.
We now have a set of possible components for a seasonal model — T,

S, /. There are several ways in which these components could be put
118
together to form a model for a set of seasonal data:

T+S+I
TxSxI
TxS + I
The first model above is called the additive model. If xr-;- is the
observation for season j in period i, then the model is formally
x = +
ij Tij + Sj hj
The second model above is called the pure multiplicative model:

Xu = Tif Sj Iij
The third example is also a form of multiplicative model, but it has an
additive irregular component. It is usual to say that a model is of an
additive or multiplicative type according to the way the seasonal
component acts on the trend/cycle terms.
What evidence there is available seems to suggest that the multi-
plicative form of model occurs most frequently in practice. Most
commercially available programmes on seasonal forecasting take this
model as their basis. It is, however, necessary to check on the form of
model most appropriate for the data on hand.
If we take logarithms of the pure multiplicative model, we get
log x,t = log Tij + log Sj + log Iij
= T}j + S} + I}j
So the logarithms of the original data can be modelled by a pure
additive model. We can thus carry out an analysis of the original data
using the multiplicative model or of the transformed data using the
additive model. The distribution of Itj often tends to be highly skew. A
property of the log transformation is that, if this is the case, log In-
tends to be much less skew and may even be approximately normally
distributed. In cases where a large irregular component occurs this
symmetry of distribution has the advantage in forecasting that
symmetric confidence limits can be given for the forecast of log x. The
occurrence of observations in the extreme tail of the distribution can
lead to complicated forms of distortion in estimates of the seasonal
component when untransformed data is used. The main consideration
in favour of using the multiplicative model in its original form is that its
various components have direct meaning in terms of the actual data,
whereas this is not the case when we have to transform the data first.
8.1.2 Moving averages

In the study of the models considered in this chapter frequent use is
made of the idea of a moving average. We therefore examine this as a
119
technique before looking at its application. In section 4.2 we

introduced the idea of a moving average to estimate a local mean. Here
we will use the same idea as a technique for manipulating data. The
smoothing produced by application of the moving average helps make
clearer the underlying variation. Table 4.4 illustrates this point. The
original data show an up-and-down movement that tends to hide
somewhat the dip in the middle of the year. The moving averages,
however, show this feature more clearly. Notice two features in the
table:
(a) There are only eight moving averages calculated from the twelve
observations. In general, if we calculate a fc-valued moving average
from m observations, there will be m — k + 1 moving averages when
k is even and m — k moving averages when k is odd.
(b) If we are to place the moving averages in time, it is reasonable to
place them at the middle of the interval covered. In the case where
k is even this corresponds to a point in time halfway between
two of the times at which data are obtained. Thus if 1 to 12 are the
months January to December the first four-month moving average
would be halfway between February and March, which may not be
very clear in interpretation. A way out of this is to average the
adjacent pair of four-month moving averages. The resultant
quantity is called a centred moving average. If we do this for the
general simple moving average, we finish with m — k centred
moving averages.
The importance of the moving average stems from the properties of
the series obtained when it is applied to data from certain models. As a
convenient notation we will denote by { Mx } the series obtained by
calculating moving averages from a series, { x}. The main properties of
{Mx} are summarized in Table 8.1. Let us comment on these parts
Table 8.1 Effects of application of a moving average M of K

terms to the following:
(a) An independent random sequence {et],E(e) = 0, Var(e) = o2

(i) Expectation, E(Me) = 0
(approximately Me = 0)
(ii) Variance, Var(Me) = o2/K
(hi) Autocorrelation, p(Met, Me t+s) = l-\s\/K (s < K)
(b) A linear trend model T, ot. + fit (M centred at t)

MT = T
(c) Pure seasonal model S of K seasons

MS = constant
120
separately:
(a) If applied to random data the moving average M reduces the
variance but leaves the mean unaltered. Thus it can be used to
smooth out the random element in data. Unfortunately, a side
effect of this smoothing is a high autocorrelation between values of
the moving averages. Thus if K = 12, the correlation between
adjacent values is 11/12, between next but one values is 10/12, and
so on. High autocorrelations such as these show themselves as a
smooth wandering movement in the moving averages, which can be
misinterpreted as a genuine oscillation in the series. Appendix A
gives a brief discussion of the idea of autocorrelation.
A further aspect of this effect is that the smooth ‘oscillation’
obtained can be emphasized even more if the sequence of moving
averages is again smoothed using another moving average. There is,
in fact, a theorem due to Slutsky, 1937, that shows that, under
certain conditions, repeated application of moving averages can
eventually lead to a sequence that follows a pure sine wave. This
result is the exact opposite of what one is usually trying to do
when using moving averages. However, this is a ‘long run’ result and
is mentioned simply as a caution and a reminder that one needs to
study one’s methods carefully before assuming that all is well. In
circumstances where one is in doubt about effects such as the
above, it is advisable to try out one’s methods on simulated data
which have known properties and then examine the results.
(b) As is intuitively reasonable, if M is applied to a perfect trend
T = a + j3£, the value taken by MT is the value of T at the centre of
the moving average. Thus apart from losing the end values M leaves
trend data unaltered. If our trend has superimposed error, the same
procedure is still reasonable; M simply smooths out this error.
There is, however, another valuable way of looking at this. Suppose
we fitted the model a + /3t by least squares to just the K
observations used in M. The value of the fitted line at the central
point will in fact be the moving average. So the moving average
operation corresponds to fitting the trend to a moving section of
data and finding the mid-value of the fitted curve. This can clearly
be extended by fitting not just linear trends but quadratics, cubics,
etc. For example, if we fitted either a quadratic or cubic to a
moving section of five observations the fitted value at the middle
would be
7~ (—3JC_2 +12x+ 17x0 + 12^! —3X2 )

35
using an obvious notation. Thus the value is of moving average
form but with weights attached to the observations. Such weighted
121
moving averages have a wide range of applications in studying

seasonal data and in other statistical applications. See Kendall
(1973) and Kendall and Stuart (1966) for a detailed discussion of
such moving averages. The general conclusion for our purpose is that
even if T is not a linear trend we may still be able to find a suitable
weighted moving average so that MT—T.
(c) If a moving average with K terms is applied to seasonal data with K
‘seasons’ in the year, each season will occur once in each
calculation of the moving average. Thus, for example, centring M
on, say, April will give the same result, for a series with a perfect
twelve-month seasonal, as centring it on January when K = 12, so a
twelve-month moving average will eliminate a seasonal fluctuation
of length 12. Thus we can write MS = a constant. We will usually
adjust this constant to be zero or one. If this moving average is
applied to data showing oscillations that take place in much less
than K observations, it will tend to treat them in the same way it
treats random fluctuations and smooth them out. If it is applied
where the period of oscillation is much greater than K, the moving
average will oscillate in the same fashion. If we consider the moving
average centred at a peak of such a series, it will contain
observations less than the peak value but none greater. Thus the
moving average will be lower than the peak value; the more values
in M the lower it will be. The converse will happen on the troughs;
hence, though the moving average will oscillate with the original
data, it will not have as high an amplitude or, putting it another
way, it will not penetrate into the peaks and troughs. One way to
improve the penetration is to use a weighted moving average where
more weight is put on the central observations.
8.1.3 Looking at seasonal data

The analysis of seasonal data has always depended very largely on
subjective judgements: initial judgements about the type of model to
try, judgements about how often to apply certain smoothing and
recalculation operations and judgements towards the end of one’s
calculations as to how good the assumptions were and whether any
slow variations are occurring and should be allowed for. The intro-
duction of more mathematical approaches to seasonal forecasting is
reducing and changing the areas of judgement for those who use these
methods. However, these approaches do not eliminate the role of
judgement. The best basis for any judgement is a careful study of the
data and the best way to start this is by an investigation based on
graphics and some simple calculations. Figure 8.2 shows various ways of
looking at a set of four years of quarterly data. Consider the various
plots in turn.
122
Quarter I 2 3 4 1 2 3 4 12 3 4 12 3 4
Data 6 II 17 10 4 10 18 6 I 9 17 5 7 10 16 7
( b ) Tier chart
(c) Centred four - term moving average, x

123
(e) Average seasonal ratio's
(f ) Rough residuals data - (centred moving average x average seasonal ratio)
( g ) Tier chart of residuals

3r •
0 —
I* 4* Quarter
-I -
Figure 8.2 Illustrative plots for looking at seasonal data

124
(a) This is a simple plot of the data. Joining the points by line
segments often helps to clarify the seasonal pattern.
(b) A tier chart helps to clarify the seasonal pattern and the extent of
its variation from year to year.
(c) A centred moving average will show up any major trends. In the
example there is no clear trend and the magnitude of the changes
are small, suggesting that the level of the series is remaining fairly
steady. The very nature of the moving average will produce a fairly
smooth curve with a natural wandering movement even when the
underlying mean is constant, such as in this example.
(d) A simple seasonal ratio, detailed calculations for which are given in
later sections, is plotted against time. Any change in amplitude or
in the form of the seasonal pattern, such as illustrated in Figure
8.1(c) and (d), will show up in this plot.
(e) These seasonal ratios are averaged to give the estimated pattern of
the seasonal variation.
(f) The irregular component is estimated by calculating the residuals.
We have done this here by subtracting from the data the smoothed
values, found by multiplying the central moving average by the
rough seasonal ratio from (e). The plot of these rough residuals
against time will indicate if the error variance is constant; with
enough data a histogram of the residual distribution could be
produced.
(g) If our rough analysis has failed to cope with the seasonal variation,
then the residuals will still contain a seasonal element. This will
show up when the residuals are plotted on a tier chart. No such
structure is evident from (g).
8.2 Seasonal index methods

8.2.1 Global non-trending models
The methods in this and the following three sections are all based on
the idea of associating with each ‘season’ an index S. Thus Sj measures in
some sense the ‘seasonal effect’ of season ; on observations obtained
during that time. We consider first, for ease of presentation, the
situation where there is no trend and the seasonal pattern can be
assumed to be constant and stable.
Let Xq represent an observation made in ‘period’ i, which might be
year i or week i, and ‘season’;', which might be the winter quarter; = 4
(of ; = 1, 2, 3, 4) or the month of December, ; = 12 (of ; = 1,
2, . . . , 12), or possibly a Monday in a working week, ; = 1 (of j = 1,
2, . . . , 5). Considering a model with additive random variation ez/-,
defined in our usual fashion, we could write
—
Xjj — fj.jj €ij * (; 1? 2, . . . , r, i 1, 2, . . . , t)
125
where ju/;- is the ‘seasonal mean’ for season j in period i. If the seasonal
mean is the same or approximately the same from period to period, we
could write this as
%ij ~ M/ îj U ~ 2, . . . , i — 1, 2, . . . , t)
For each season j this corresponds exactly to the constant mean model
of Chapter 4. Hence we can use the formulae of that chapter to provide
estimates of /x;-, here denoted by jut f. Assuming a global model, we have
t
Mt,j ~ ^ %ij/t ~ X.j
i- 1
The layout of the data and the calculation of pLt )- for the global model
is illustrated in Table 8.2. The forecast of xt+ x j is simply pLtj.
Table 8.2 Constant seasonal structures — calculations

for global model
Season j
1 2 r
1 *i I *12 Xi r
2 x
2 1 *2 2 *2 r
Period
t Xtl x
t2 Xfy
Totals X 1 *2 X r
Hi X 1 *.2 X r
It is clear that the forecasts of each juj are based only on data from
that period, so we are really doing separate forecasting exercises using
the constant mean model for r different sets of data. For the global
model this is obviously the correct thing to do.
Suppose that we wish to refer to the seasonal means \uy as being
relative to a constant mean (JL for the whole time of observation. We
could do this in two particularly simple ways, either by using an
additive form
M/ = li + 0/
or in a multiplicative form
M; = ndj
So in our notation the trend term T is now a constant mean p and the
126
seasonal term Sj is given two expressions, 0;- and Oj, according to

whether our model is additive or multiplicative. In both cases we would
want p to be the average of the p;- values, i.e.
1
v fij
p =— 2 r 1
/= i
and therefore we must have conditions on 0;- and Oj such that
2 0; = 0 and 2 0,- = r
7=1 7=1
Table 8.3 illustrates these formulae with some numerical values. It

should be emphasized at this point that on the basis of the global model
these formulae are just different ways of writing the same thing: p;- is
the fundamental quantity p, 07- and Oj are artificially introduced
parameters called seasonal indices. When 07- is used, we say that we are
using an additive form of model. When Oj is used, it is called the
multiplicative form. We have introduced these artificial forms here as at
this stage they are easily understood. When we come to more
complicated seasonal models, we will find that these two forms are
fundamentally different and not just different ways of writing the same
thing. Having introduced the parameters p, 07 and Oj, we now ask how
to evaluate them. To do this we first write our model in its two new
forms:
xtj = P + 0/ + etj
and
Xij — llQ) "t" €(j
If we average over j for each /, we obtain

r
1
xt. = P +” 2 0; + ef#
r
7=1
1 ^
Xi = p X — 2 Oj +
r j= i ’
which by the conditions on 0; and Oj gives

xim = P + eu
in both cases. Taking expectations gives
li = E(Xj )
so p is the underlying mean for the average of the data from any
period i.
Let us assume for simplicity that there are exactly tr observations, as
G
03 O
03
is—1 O O O rH
S T—I rH
cj CD
03 co o ^ o'
Q 1-1
1
> 00
o CO O <M o'
tH
1
-4-i CM
03 CM O CM rH
O T—1 T—1
-l-s U0
Qu
cn 03 CO O CO T-H
S-i cn T—1 T—1
03
03
£
o3 eii) uo
J-i
03 G WOiOH
a < T—1 TH
G
o
C/3 LO
o3
03
C/3
3
l
“5
lO O lO H
T-H T-H
«4-l
O
_a>
03 CO
'a G
£ G
l—
3
COOCOri
T-H rH
03
X
w
co T-H
>>
00 03
i—i T-H O TH T—1
03 T-H T-H
3
03
J-I 03
a 03 O T-H O
<3 ^ 1
3-4 C—
03
O' O CO o
^ 1
CD
£>
03 CDOrJO
&H
^ i
UO
•
G*
o3 uO O u0 O
'”5 tH
1
3-1
03
-P
03
G
G
o3
3-H
cO
PH G. G.-0-<G>
127
128
in Table 8.2. The average over all seasons, xL (i = 1, . . . , t), has

expectation p for all periods; thus the average of the x t over all i will
provide the estimate for p obtained at time f,
t
fit = x.. = 2 Xj. /£
/= i
which is the ‘grand average’ of all our data. If instead of averaging over
seasons we average over periods for each season, we obtain fitj once
again. In terms of our two forms of model this is
A ,j = n + <l>j + i..f
and
At,i = +e.j
where e denotes the average of e,-7- over i. It is now natural to estimate
(j)j and 6j by
Sj = frt.j - i*t
and
Qj = fit, jib
We thus have reasonable ways of breaking our estimates of into
components jd and 0; or p and Oj. A little calculation will show that
£ 07- = o
/= i
so the additive seasonal indices are positive and negative numbers

indicating deviations above and below the mean. Also,
r A
2 0y = r
7=1
so the multiplicative seasonal indices are numbers above and below one,
indicating deviations above and below the mean. The forecast of the
value of any further observation xT j- in season) is
xTj = At + 0/
or, equivalently,
XT,i = A»A;
The structure of the above calculation is presented in Table 8.4. If
the assumption of exactly tr observations does not hold, it is clearly no
use basing a period average on less than r seasons’ data since not all
values of 0; or Oj will be included. In this case jlt could be based on an
129
Table 8.4 Global and local fitting of models in section 8.2
Season
Period Exponential
1 2 r average weights
1 *11 *12 *1 r *i. at~1

2 *2 i *2 2 *2 r *2. af~2
3 *3 1 *32 *3 r *3. a{~3
Period
t-1 *(f-l)l *(f-1)2 *(f —1)/* *(f — 1). a
t *f 1 *f2 Xff xt. 1
Column averages
or weighted
averages fit, l fit,2 fit,r fit
0/ fit, l - fit fit,2 ~~ fit Mt, r fit
9/ fit, l / fit ^7,2/Mr fit, rl Mf
average of only those periods which contained a full complement of

seasons or, alternatively, on an average of the r seasonal averages. These
two estimates of p will not be the same, as the latter will use all the
data and the former only part of the data.
8.2.2 Local non-trending models
We now turn to the situation where we feel the need to consider

fitting our model locally rather than globally. In this context this means
that we will consider methods that allow for slow changes in the
seasonal indexes as new data arrive. Remembering that our basic model
is
*// M/ + îj 0 1, 2, ...,r, i 1, 2 ,...,£)
which is the ordinary constant mean model for each j, an estimate of

during the most recent period t, using exponential weights, is
t
at - i
i= 1
This is exactly as discussed in Chapter 4.

To estimate the underlying mean p, a number of estimates could be
considered. Of these the weighted mean of the ordinary ‘yearly’ means
will be used, i.e.
130
This form has the advantage that if we define 0; and 0as before, by
h = Ar,/ - Ar
and
0,- = At,,7 At
then all the results of section 8.2.1 still hold for the local model. We
also have, for large t, the recurrence relations
+ a
A,,/ = (1 — a)xf; Ut- l j
and
At = (1 — + aAt- l
The above equations provide an ‘exact’ type of solution together with a
recurrence form for large £.
Let us consider now the estimation of the parameters ju, 0y and 07- in
a purely intuitive fashion and with the object of expressing our
estimates in the form of recurrence relations, such that jlt can be
updated more frequently than once a year. We thus want something in
the form
New estimate = constant x new information + constant x old estimate
We have already an expression for [itj in this form. If we wish to get

this form for the seasonal indexes 0y and 07, we could subtract or divide
by the latest estimate, jlt, of the underlying mean. Subtracting A? from
both sides of the expression for Ar,/ gives
Af j—fit = (1 — a)(xtj — A,) + o(At - 1,/ — A,)

To make the use of this expression a practical proposition, we must
rewrite it in an approximate form, namely
Af,, — Af = (1 — a)(xt/ — A,) + a(Af- 1,/ — Af- 1)

where in the last term we have had to subtract \ut_ x instead of jOlt to
correspond to the fact that at time t — 1 the estimate of the underlying
mean was A?- I • In terms of 0 this is
A A
0f,/ ~ (1 a)(xti M?) ^0?— I ,/

A
Similarly, for 6 we would have
Of ,i = (1 —a)(Xtj/M + O0f_ 1,/

Using these we can update estimates of the seasonal factors 0 and 6
provided ptt is known. We now want to replace jlt by a form in which
we can update the estimate of ju using each observation as it is obtained.
Let us denote such an estimate made in season j of period t by A(?,/)• A
131
reasonable recurrence relation for A(?,/) would take the form

= —
(1 b) (latest information about ju) + bj±(t ;_ i)
where b is again a constant 0 j. The recurrence equations thus become
ht.i) = (1 -b)(xtj — 0,_1 J-) + bA(f,/-i)

and
£<r,/) = (i_ b)xti/et_, tl + 6A(f,/-1 >

where in both cases we have used estimates of 0 and 0 obtained at time
t — 1, since the estimates at time t cannot be obtained until /}(*,/) is
known. Having obtained a recurrence relation for /}(?,/) we may use this
in place of Ar in the recurrence relations for 0y and Ay. The quan-
tities xtj — 4>t- i,/ and xtfj/dt_ ij are termed the seasonally adjusted
observations, since they are essentially our estimates of what the
observations would have been had there been no seasonal effects. As is
usual with recurrence relations, we can re-express them in error-
correction form. Thus denoting the one-step-ahead forecast error at
season j of period t by et y, we have
A
e x +
t,j ~ t,j (A(t,j- i) fit- i,/)
or
e
t,j ~ xt,j (Ai)®t— i,/)
Then we may rewrite both the recurrence equations as
= +
£(*,/) ht,j- i) (l—b)et>f
The sequence of calculations is as indicated in Figure 8.3.

As on previous occasions we will postpone a discussion of the choice
of forecasting parameters a and b until a later chapter. However, it
should be noted that b should be very close to one, otherwise the
natural variation in A(t,j) will distort the estimates 0 and 6 and the
process of forecasting will tend to lead to erratic results. It might in fact
be safest to use previous methods discussed for calculating jdt at the end
of each period and to use this value as the initial value in calculating the
A(f+ i j) values for the r seasons in the following period.
In the global model situation we noted that the difference between
additive and multiplicative forms was purely arbitrary and the answers
obtained were identical. In the recurrence relations above this is no
longer the case. In a local situation the underlying mean will wander.
There is thus a real difference between the situation where an additive
132
Error from
previous forecast
Update
mean estimate
New observation
*t.j + l
Update index for

season j
Forecast for next

period j + I
Forecast for any

period k
Figure 8.3 Recursive calculation of the mean and seasonal factors
seasonal 0 adds the same amount on to fjtt, whatever its value, and the
multiplicative situation where an increase in jit will increase the total
variation in the seasonal pattern represented by 0/xf.
In the above we have assumed that the seasonal element, S = 6 or
S = (j), wanders in a slow and irregular fashion. There may, however, be
systematic change taking place in S. This possibility should be
examined by looking at the variation of S with time. The quantities
A ^
— X(j fif and Sij — Xij/jlf
give, for additive and multiplicative models respectively, an estimate of

the value of S at time f, which as before in terms of periods and seasons
is (ij). A plot of Sij as a time series will give a general indication of any
systematic change taking place. More detail can be found by looking,
separately, for each season; how Sj,- is changing with i. If there is a clear
increase or decrease in Sjj with i for any reason j, a forecast of future
values can be obtained by extrapolating the trend.
8.2.3 Models with trend
In the previous sections it helped the presentation to distinguish

between global and local approaches to fitting and extrapolating the
seasonal model. When we turn now to seasonal models with a trend
133
component, the clarity of distinction gets lost when seasonal index

methods are used. Basically the methods are designed for use in
situations where the trend component is unlikely to be a simple
mathematical trend and thus the whole approach tends to look at the
situation in ‘local’ terms. There are a number of points in the procedure
where, if the situation is thought to be stable, an ordinary average
might be used or a mathematical curve fitted. If this stability is
doubted, exponential smoothing might be used instead. The general
procedures, however, are common and it is these that we will study.
Where a trend is present any attempt to derive seasonal components
using the methods of the previous section will fail because the seasonals
would contain contributions from the trend. To find the seasonals we
must first find the trend and remove its contribution. There are two
elementary methods of doing this. One is to find an empirical trend
curve by a suitable smoothing method. The other is to fit a
mathematical trend curve. Let us examine these briefly in turn.
(a) Fitting an empirical trend curve
The basic method of obtaining a trend for data of seasonal period r is

to use a simple moving average of r terms. To see the justification for
this consider the additive model
x
ij = Tn + Sj + lij
Applying the moving average operation centred at time ij gives
Mxfj — MTij + MSj + Mlij

Using the results of the section on moving averages
MT^ = Tn
MSj = constant = 0 by choice

=
MItj 0 (approximately)
Hence
Mxij - Tij
and the sequence of values given by the moving averages approximately
follows the underlying trend at the centres of the moving averages. If
the model is multiplicative, we will get M(TijSj), which does not
conveniently simplify unless T/y is nearly constant over the time
spanned by the moving average. We might in this case convert to an
additive model by taking logs of the data and then converting back the
estimated value of T/y obtained.
134
(b) Fitting a mathematicab trend curve
Before we can fit a trend curve we first need to get rid, at least
approximately, of the seasonal component. This can be done most
simply by using a period average in which the seasonal component is
summed out, as it is in the moving average. These averages can now be
fitted by an appropriate mathematical curve using the method of least
squares or any other appropriate method, we have already discussed
fitting straight lines in Chapter 5 and we will discuss fitting non-linear
curves in a later chapter. However, there is a complication that requires
a brief mention, and that is that the period averages do not lie on the
same curve as the original data. For example, if monthly data showing
slope |3 are used to get the averages for a sequence of years, these yearly
averages will increase by 12 j3 from year to year. However, if the
original data, the averaged data and fitted lines are plotted together,
there should be no confusion in obtaining the correct estimates.
Both the above methods are open to criticism and the choice
between them depends somewhat on how appropriate these criticisms
are in one’s own particular circumstances. The main aspects to be
considered are as follows.
(a) Empirical curves
The main criticism of these curves is that they are empirical. They do
not have a specified mathematical structure. This fact has two
unfortunate consequences. Firstly, as we do not have a structure
defined, we cannot easily extrapolate into the future to obtain our
forecasts. Secondly, the lack of a model makes almost impossible an
examination of how well the method has determined the past trend.
The method really provides one definition of what is meant by the
trend.
(b) Mathematical curves
If a very simple mathematical model such as a linear trend is used, it

is usually too simple to describe any realistic trend and will certainly
not include the cycle. However, if a much more complicated model is
used, it may well tend to pick up parts of the seasonal structure as well
as the cyclic structure.
We started this section by saying that before we could estimate the

seasonal indexes Sj we had first to find and eliminate the effect of the
trend component T?/-. We have now found methods of obtaining at least
a rough estimate Ti}- of this, so let us now continue by finding an
135
estimate Sj of Sj. Assuming still an additive model we can get a first

estimate of Sy, using observation Xq, by subtracting T/y, giving
S ij — X(j Tfj
From the model this quantity will be seen to be approximately
Sj + Iij
To estimate Sj then we must now reduce the effect of 7Z/- by averaging
the above quantities over z,
A A
Sj - average of S/;- (j = 1, . . . , r)
i
The sequence of steps is thus:
A
(a) Estimate Tiy by a moving average or fitting procedure, giving Tzy.

(b) Subtract these from the original data, xij ~ Tij.
(c) Average over all periods for each season.
We now have estimates of both Tz/- and Sj so we might go ahead and

produce a forecast. However, our methods for finding the estimate of
Tfj left a lot to be desired, so in practice we seek to improve our
estimate first, As we now have an estimate Sj of Sj we can seek to
remove the seasonal effect by division or subtraction to get seasonally
adjusted data
xij X{j/Sj or Xij ~ Xij Sj
If the seasonal was found in the previous investigations to vary in some

fashion, then Sj here would be replaced by our best estimate of S at
time ij.
The aim of the next few pages is to discuss again the estimation of
Tij. However, we now have seasonally adjusted data as the basis for our
calculations rather than the original data. With the influence of Sj at
least approximately removed, the object is now to seek to estimate Tz/-
by reducing the effect of the irregular component 7zy. This may be done
by fitting by eye, by fitting some empirical trend, using some type of
moving average technique, or by fitting a mathematical trend curve.
Considering the latter two in turn.
(a) Moving average techniques
A natural intuitive method of reducing the effect of the irregular

component is to smooth it out using a moving average of some form. A
traditional way of doing this is to use a weighted moving average to fit a
polynomial to a section of data. For example, suppose we choose a
section of five observations, called il2, x!_ 1? XQ, X[ and x^, and fit a
linear trend by least squares. The simple moving average is just the
136
average of the five observations. Now we have seen before that the
trend line goes through the mean of the observations, so the value of
the fitted line at the time corresponding to x'0 is
—
(x'_ 2 + X- l + XQ + x[ + x'2 )
5
often denoted by
“ [!> 1,1]
o
making use of the symmetry and listing only the coefficients of x'. If
we had fitted either a quadratic or a cubic, the value of the fitted curve
at the midpoints would have been
— ( — 3 x'_ 2 + 12 x'_ ! + 17 x'0 + 12 x[ — 3 x2)

35
or
— [-3, 12,17]
35
Thus as the above five-pointed weight moving average is moved

through the data, it effectively fits a quadratic or cubic (the result is the
same) to the successive sections of data and gives the midpoints of the
fitted curves.
One disadvantage of using the various types of moving average is that
they estimate only the trend at the middle of the section used. When a
2 m + 1 moving average is moved through a set of data, no trend value
is estimated for the first and last m values. If one is using a weighted
moving average to fit a curve by least squares, then one can give the
values of the fitted curve over the m end values and indeed extrapolate
past these. A danger here is that only a relatively small number of
observations are used in the estimation, so that extrapolation of, say, a
cubic to the end of the moving average can lead to estimates with large
variances. An approach sometimes used, though of debatable logic from
the forecasting viewpoint, is to forge extra m observations at either end
of the data being used. An inappropriate choice of forged values will
affect 2 m of the moving averages and possibly lead to incorrect
interpretations and forecasts.
An alternative approach here is to fit a trend curve using the two-way
exponential weights (see Appendix C). This enables one to obtain
estimates at all points of the data, for the maximum weights can be
placed successively at each point in the data.
The great disadvantage of this approach is that it produces an
137
empirical trend curve over past data which has no natural way of
extrapolating into the future. It does, however, provide a means of
analysing past data as a preliminary to forecasting.
(b) Fitting a curve

As an alternative to the above we may seek to fit, perhaps by
discounted least squares, a simple mathematical curve to the seasonally
adjusted sequence. For forecasting purposes this is a sensible procedure
as we are now provided with a means of extrapolating the trend into
the future. Consider the two simplest cases of local constant mean and
trend models. In the former we could update the mean now given by
Tij using exponential smoothing. Thus
Ttj =atiJ_i + (1 —a)x'u
A
where Tz 7 _ x denotes the last estimated trend value, ignoring whether it

was in period i or not, and x'a denotes the latest seasonally adjusted
observation. At this stage only Si- \ j is known, so
Xu = xif/Si- i j, for the multiplicative (x) model
or
A
xijt ~ 0C{j — Si- i j9 for the additive (+) model
A
The natural formula for obtaining the Siyj is that given at the end of the
previous section, namely
su = bs, J + (1 - b )Sjj
where the simplest form for Siy- is
A
Sij = Xij/Tij, for the multiplicative model

or
A
=
Sij Xij — T{j, for the additive model
The methods of (a) above and of the previous sections can be used
on an initial set of data to provide the values from which to start the
recurrence relations. The sequence of the recurrence relation is given in
Figure 8.4(a).
The forecast of a future value xrs will be
xrs = T^ + SiS (+ model) or TijSis (x model)
where Tz;- is the latest trend estimate and Sis is the latest seasonal factor
for season s. If a study of past data can give good starting values for the
recurrence relations, it is advisable to keep the two forecasting
parameters a and b quite close to 1.0 otherwise such a set of equations
138
(a) Locally constant mean form
( b) Local trend form
Figure 8.4 Sequence of calculations for recursive forecasting
can easily lead to wildly varying values. If there is evidence of a change

in mean and one wishes to pick this up rapidly, a would be reduced but
not b — conversely, if a large change in seasonal pattern is suspected.
If the model is a local linear trend, we would need recurrence
relations for the current level of the trend Tfy and for the slope ft-y. The
natural relations to use are those of Chapter 5; thus
+ _
Tij = I + ft,/- I ) (! a)x'u
Ay = eft,/- i + (1 — C)(T,7 - Tjj_ ,)

A
The recurrence relation for S/y will be as before. A forecast of an

observation xrs which is h units into the future is
Xrs =
{Tij + hfiij) + Sis
The sequence of calculations here is given in Figure 8.4(b) and is often
called the Holt—Winter’s method.
A slight modification of this procedure is based on the observation
that x'ij, the seasonally adjusted observation, is based on the use of the
last seasonal factor St_ { y. Having gone once through the sequence of
139
calculations in Figure 8.4(a) and (b) we can use the new value to
revise x\j and iterate the calculation to obtain what are, intuitively,
better values for Stj and fty.
The literature on seasonal methods gives extensive study to ways of
cycling round these various calculations seeking at each stage to
improve these estimates. Durbin and Murphy (1975) and Harrison
(1965) discuss aspects of these calculations in detail. Basically, the
modifications are concerned with:
(a) reducing the effects of extreme observations, possibly caused by
known factors;
(b) adjusting the seasonal estimates so that they satisfy the conditions
that XS = 0 or XS = 12;
(c) introducing more sophisticated forms of moving average to deal
with particular features of the data.
8.3 Fourier methods

Over the years many criticisms have been made of the iterative
approach described in the previous sections, particularly where it is
applied to the multiplicative model. For example, Lovell (1963) and
Jorgenson (1967) have stated criteria that a good method for fitting
seasonal models should satisfy; in general, the methods discussed above
satisfy neither of these sets of criteria.
Particularly relevant criticisms from the forecasting viewpoint are
(a) the lack of any estimates of the error variances and hence of
prediction errors. The only way round this is to estimate these by
applying the methods to past data.
(b) the lack, in some formulations of the approach, of a mathematical
model that can be extrapolated.
In spite of these criticisms we have examined these methods at some
length as they do provide a robust approach that has led to several
simple methods of forecasting. Our study has also enabled us to
introduce a number of ideas of importance in other approaches to the
analysis of seasonal models. The aim of this section and the following
section is to give a brief survey of some of the alternative approaches.
The essence of most of the recent methods of seasonal analysis is to
express the model in additive form and to apply to such a linear form
of model a standard method of estimation. Thus the basic model for
the observations, or their logarithms, is
x =
ij Tij + Sij + Itj
where each of the three parts can be modelled in some sensible way.
For example, the trend might be modelled by a polynomial in t (see,
140
for example, Duvall, 1966). The irregular component might be

modelled by a set of independent normal or log-normal variables or by
some simple stochastic process. The main focus of recent work
however, has been to model the seasonal component using modifica-
tions of methods that date back to the last century.
The basis of the methods is to fit the seasonal component with a
regression model in which the regressor variables are periodic functions
of time. Suppose, for example, that we have monthly data with twelve
seasonal indices, Sl9 ... , S12. The methods called Fourier analysis
show that this sequence of twelve points lie on the curve
6
277 kj
S; = 1 + 2 cos ——- + vk sin
k= 1 b
This is a regression curve with regression coefficients uk and vk given by
1 L1 „ 27r kj
uk — 2J Sj cos (k = 1, . . . , 6)
6 /= I 6
and
1
v
2 s • 277 kj
c Sln (k - 1,. . . , 5)
vk = T i —T~
6 j=\ 6
= 0
Thus we have converted from one set of twelve indices with the
additive condition
(t -«)
to another equivalent set of eleven regression coefficents. Thus at this
level there is no great merit in introducing such a Fourier model. The
advantage of these models comes from the practical observation that
one usually obtains a good fit without having to use all twelve terms in
the regression model. Any pair of sin and cos terms, say,
u
27r kj . 2Tr k
k cos —- + vk sin —— ]
represents a pure oscillation with amplitude
o-k = \/(u2k + vl)
It may be observed, for example, that ax and a2 are large but a3. . . a6
are relatively small. The procedure then is to use the model
27Tj 27Tj 477 7 4777

S, = 1 + ux COS
6
+ vx sin
6
+ cos
6
+ vo sin
6
141
The coefficients u1 v x, u2 and v2 can be found by fitting this as a linear

regression model to the latest seasonal indexes obtained by the methods
of the previous section. Such a method of replacing the usually rather
irregular Sj values by a smoother form of seasonal variation is referred
to as ‘Fourier smoothing’ (Harrison, 1965). In general this model would
have to be fitted each time a new observation and estimate of S was
A
obtained. However, if each new S is obtained by a recurrence relation

from the previous value for that period, as discussed in the previous
section, then recurrence relations can be obtained for the coefficients
ux, vx, etc., and for the smoothed seasonals themselves.
Thus we have
Sij = Si_1J + (l-b)(Sij-Si_lJ)

which is the ‘error correction’ form of the recurrence relation of the
previous section. Let §jtt denote the smoothed seasonal obtained at
time t for season j. Then
Sk.f = Sk,t- i + constant x (S,7 — S/>f_ !)

Note that as each new Sf/- is obtained we fit the curve over all seasons.
Thus we update all seasons, fe, on the basis of the new information
about season j, S^- and the previous information about season k,
Skt_ x. The ‘constant’ does in fact depend on j and k but not on the
data. It can be written as
2(1 - b) ^ 2 r(k — 1)
2 COS
n r n
where n is the number of seasons and the summation is over the

number of sin, cos pairs used in the Fourier model; in our example this
is two.
A number of variations on this theme are discussed in the paper by
Harrison. These methods have a number of advantages. They combine a
mathematical model for the seasonal component with some of the
standard methods for dealing with trends. This enables us in effect to
reduce the number of parameters that we are using. By use of different
smoothing constants for the seasonal and the trend fitting the methods
can allow for changing trends in a fixed seasonal pattern and vice versa.
The use of the fitted Fourier terms that are dependent on all the
seasonals also fits in with one’s intuitive feeling that in most situations
the seasonal components are not independent of each other with one
able to vary without effecting those on either side. The Fourier
approach links the seasonals together in an intuitively reasonable
fashion.
A natural extension of the above approach is to modify it to deal
with the possibility of changing seasonal patterns. Whether this is the
case can be investigated by plots of the u and v values and amplitudes as
142
they are recalculated to allow for more data. One natural extension is
to make u and v linear functions of time; alternatively, the whole
seasonal expression can be multiplied by a time-varying term. The work
by Nettheim (1964) and Duval (1966) deals with fitting these forms of
model. In such applications we usually have a linear regression model
and so the classical approach is to fit by least squares. We have already
seen that discounted least squares with exponential weights ar reduces
to ordinary least squares when a = 1. We will therefore look at the
fitting by discounted least squares rather than the special case of
ordinary least squares. If we seek to reduce our problem to a linear
regression form of model, it is natural to include the trend in the model
and not just the seasonal, as we did above. Examples of linear models of
use in seasonal forecasting are
27r t 27T t
X, = /So + Pi t + p2 sin + (3 3 cos + et
n n
2i7
Xt = Po + (P2 + 72 Osin — VP 2 / 5 -
n n
277 t 27Tt . 477 t 477 t
xt=Po+ p2 sin + p3 cos + p4 sin cos + et
n n nn n
In the first example an additive linear trend is included with the simple
seasonal. This has clear advantages in terms of forecasting when the
situation is fairly stable. If, however, this is not the case and
independent changes can occur in the trend levels and the seasonal
patterns of the data, then the previous methods of separately handling
T and S are likely to give better results. When we fit by discounted least
squares the smoothing constant a is clearly applied to the whole model,
and having different smoothing constants for the trend and seasonal
parts is clearly ruled out. In the second example the amplitude and
position of the simple seasonal pattern is changing in time. In the last
example a more complicated seasonal pattern is allowed for by
including more Fourier terms. As a means of indicating some of the
procedures used in forecasting from such models, let us consider in a
little more detail how we could locally fit a simple linear model with a
trend and seasonal component. Suppose the fitted model takes the
form
7 T . .
277 t 277 t
xt = be + b 1 + 1 sin + o3 cos + et
n n
where et is the residual and 60, bx, b2 and b3 are discounted least
squares coefficients chosen to minimize
5=2 aret_r
r= o
143
This is simply a linear regression model with 1, f, sin 27xt/n and cos
2nt/n acting as regressor variables. As usual the values of b0,bl9 b2 and
b3 are obtained by solving the set of normal equations. With a slight
alteration of the notation of section 6.2, the four normal equations are
60S(1,1) + &iS(l,t) + b2S(l,s) + b3S(l,c) = S(l, x)

b0S(t, 1) + b1S(t,t) + b2S(t,s) + b3S(t,c) = S(t,x)
b0S(s, 1) + bxS(s,t) + b2S{s,s) + b3S(s,c) = S(s,x)
b0S{c, 1) + blS(c,t) + b2S(c,s) + b3S{c,c) = S(cjc)
where S( ) is defined by the formula

t
S(z,v)= 2 arzt_rvt_T
r= 0
Thus, for example,
S(s,t)= 2 ar sin MLZl) (t - r)

r=0 n
and
27r(t — r)
5(c,x) = 2 arxt~r cos
r=0 n
In practice we have a large, but straightforward, exercise in calculation
to evaluate all the coefficients, S( ), and to solve the normal equations.
The forecast is obtained as usual by setting the future error term at zero
and substituting the future time t + h in the fitted model. Hence
b0 + bl(t + h)
The above linear regression approach can clearly be extended to include

more terms in the model. Often more terms from the Fourier series are
included, in particular one would usually include sin Air t/n, cos 47r t/n
and possibly sin 67r t/n, cos 6?r t/n. It is also sometimes useful to
include terms that allow for a trend in the amplitude of the oscillations,
e.g. t sin 27T t/n and t cos 27T t/n.
The above formal solution of the normal equations involves a
considerable amount of computation, when it has to be repeated each
time a new observation is obtained, the work can get too extensive for
practical use. As a consequence of this, techniques have been
introduced that reduce the amount of calculation substantially. There
are three main elements in these techniques.
(a) The introduction and evaluation of new trigonometrical functions

required for updating the S quantities can be avoided by the use of
144
trigonometrical relations. Thus when xt+1 is introduced, we would

require to know sin 2ir(£ + 1 )/n and cos 27r(t + 1 )/n. However, these
are related to the previous values, sin 27r t/n and cos 27r f/rc, by
standard trigonometry:
. 27r(£ +1) . 2irt 2TT 2nt . 27r

sin — = sin cos — + cos sin —
n n n n n
27r(f + l) . 2irt . 2ir 2irt 2ir

cos = — sin sin — + cos cos —
n n n n n
Thus the evaluation on the computer of trigonometric functions

can be replaced by simple multiplication and addition, using the
constants sin 27r/rc and cos 2?r/n.
(b) The second technique is one that we have already made use of in
our discussion of the linear trend model, namely the constant
shifting of the origin to the present. With the trend we started with
xt = a + fit
but then rewrote the model as
xt_r = at — fir
so that 0it always referred to the mean of the trend line at the time
of the latest observation. This requires the updating of the meaning
of the parameters. Here, for example,
=
°^t+ I &t + fi
This has the advantage that the forecast for a lead time h using
estimates at and $t always takes the form
x,+h =0it+ I%h

To illustrate this let us take the mean and trend terms to be zero
and consider b2 jt and b3 tt as the fitted values for the model which
takes the latest observation time, t on the absolute scale, as the
origin. Thus the forecast corresponding to the above trend would
be
=
. 27T h . 2ir h
Xt+h b2 ,Tt sin n + 63 ’ t cos n
We will clearly wish to update the model parameters so that at time

t + 1 we will know the relation of the new parameters to the
previous values at time t. This can be done by making use of the
trigonometrical formulae used in (a) above.
Considering the underlying model, with parameters fi2 t and fi3 t
for origin at time t. If we write the underlying mean value of the
145
series for time t + h + 1, treating time t + 1 as origin, we have

. 27T h n 27T h
P 2,7+1 sm + p3>f+i cos
77 ’ 77
If, however, we write down the same value using t as the origin, it
takes the form
. 27T(/Z + 1) 27Uh + i)
0 2,7 sm + p3 f cos
77 77
Using the same formulae as in (a), this last expression can be put in
terms of sin 27T h/n and cos 27T h/n. Equating the coefficients of sin
27T h/n and cos 27T h/n in the two expressions gives
« ,t+i = Pi.t
P2 a 2TT
cos — — . —
sin 2ff
IV IV
and
2TT
P 3,7+1 COS
77
Thus we have again a simple way of updating the model

coefficients so that we can use as a model one that takes the latest
observation time as its working origin. Taking the present time t as
origin, the ‘S’ terms in the normal equations all take the form
7
St(z,u) = 2 arz_rv_r
r=o
So, for example,

7
. 27TV
St(s,t) 2 ra r sm (the two — signs giving a + )
r=0 n
and
7 27rr
r
St (c,x) 2 a xt_ r COS
r-0 77
Solving the appropriate normal equations in this form at any time t

will give fitted values b2 ,7 and b3 t •
(c) The results in (b) give a way of updating the model parameters with
shift of origin. However, in updating the fitted parameters we need
to introduce into the situation the latest observation. To deal with
this aspect we can use the fact that all the s( ) terms in the normal
equations can be updated in a simple fashion:
St+ J (z,v) = zt+ lvt+l + aSt(z,v)
This is an exact relation and can be generally used for updating.
146
This recurrence relation would always be used for updating the

terms on the right-hand side of the normal equations since these
will involve the introduction of the new observation xt+i. If we use
the present as the origin and have a large amount of data, we can
get an approximate result that simplifies calculation considerably.
In this form, from (b), we have
t+ 1
=
Sj+i(z,i') 2 arz_ rv_ r
r-0
SO
St+ 1 (z,v) = St(z,u) + at+ 1 z_ (f+ I )V_ (,+ 1)
For all the st terms on the left-hand side of the normal equations
the quantities z and v are sine or cosine functions that are
numerically less than one, or are simple powers of t\ hence for large
t the added term at+ 1 (t+1 (f +1) becomes smaller and smaller
as t increases. It can in fact be shown that for large t all S terms
on the left-hand side approach constant values. As these values do
not depend on the data, they can be worked out as a set of
numerical values Soo(z,v) as a preliminary to the forecasting
exercise.
Notice that this approach exactly parallels the methods discussed
for fitting a linear trend model. In that case discounted least
squares gave a set of normal equations which for large t led to the
simplification of the recurrence relations of the double exponential
smoothing method. If the values S«, ( ) are used in the normal
equations, from the start of the forecasting method, we obtain a
method that approximates the discounted least squares solution in
the same sense that the double exponential smoothing method does
in the linear trend situation. Further, if explicit solutions for the
fitted constants are obtained by solving the normal equations, then
these solutions, though rather complicated, can be expressed in
recurrence form. They can also, as has happened before, be
expressed in error-correction form. Suppose we denote the one-
stop-ahead error by et+ l. Then
27T.1 27T.1
e
t+ I = **+i ~ b2 t sin b3 , cos
n n
Then using the results of (b) we can show that
27T 2ir
b2,t+i = b2,t cos b3 f sin — +k2et+l
n n
27T 2TT
b3,t+\ = b2,t sin — + b3J cos—+k3et+l
n n
147
The parameters k2 and k3 controlling the amount of response are

direct functions of a. They can be found using the normal
equations, with t as origin, by replacing b2 ,t by k2 > b2 t by k3 and
St(z,v) by the value of z at the present, zero time. However, as with
Holt’s method, we may regard these results not as an approx-
imation to the discounted least squares method but as a new
method in which the user can control the two parameters k2 and
k3 to get the best results.
It should be noted that, for both methods, as the number of
model parameters increases we need to make greater use of past
data to get viable estimates. Thus the optimum value of a is likely
to get closer to the global model value of one; the values of kx and
k2 are likely to be correspondingly small.
It is hoped that the above discussion of one example has been
sufficient to give the reader enough understanding to construct the
normal equations for other models and, possibly, to get these to a form
that can provide a workable solution. The usual presentation of these
results makes use of a matrix presentation that gives a very elegant way
of studying the method. The reader is referred to two books by R. G.
Brown (1963, 1967) that deal with this approach clearly and
thoroughly.
8.4 Stochastic methods

In recent years there has been a growing interest in using stochastic
models to forecast seasonal data. The basis of this approach is to
construct models in which there is a stochastic relationship between
each observation and (a) those close to it in time and also (b) observa-
tions in the same season of previous periods. To construct such models
we make use of differences between observations in the same season of
successive periods. Thus we would define
Vrxt =xt — xt_r
Quoting an example from Box and Jenkins (1970), we might, for each
month separately, construct a model
v1 2xt = a, - dat_ i 2
where the ce values form a stochastic process. As the second stage we

now build a model for the process of a values. We might, for example,
have
V«* = — j3e,_ t
where the et values form an independent sequence. Substituting for ott

from this second model in the first and expanding, we finally have the
148
model
x( — i — 12 + X,_ , 3 = e; — 0e,_ , — det_ , 2 + 0/3e;_ , 3
This can now be used to obtain forecasts using the method discussed
for stochastic models. This model provides a very elegant form of
seasonal model using only two parameters to describe a fairly complex
situation. A somewhat similar approach is used by Grether and Nerlove
(1970), who use mixed autoregressive—moving average models indepen-
dently for the trend—cycle and the seasonal components of a classical
additive seasonal model. A detailed discussion of the methods for
identifying and fitting such models is beyond the scope of this book,
but is given in the book by Box and Jenkins (1970).
An alternative approach that leads to similar forecasting formulae to
the above is based on using a regression of xt on recent values and
values a year or so ago. Thus the model might look like
+ + e
xt = b0 +blxt_1 + b2xt_l2 &3*t-i3 t
The actual terms used in the model can be determined by the usual
methods for selecting variables. A brief discussion of this approach is
given in section 15.4.
8.5 Comparison of methods

The previous sections have looked in some detail at the problems and
methods of seasonal forecasting. The first motive for such an extensive
treatment was the fact that seasonal variation occurs frequently in
practice. The second was that it provided opportunity to discuss
forecasting using models that contain several different components
showing different types of behaviour. We have looked at three
approaches to forecasting in seasonal situations. Which method is best
in any situation is clearly a matter of practical investigation. There is
some evidence to guide one, e.g. Reid (1972), Groff (1973). If the
irregular, random, component is small, then the seasonal index methods
are probably best. If this component is large, however, it makes the
estimation of the seasonal components difficult and the Fourier
methods are likely to be better. Fourier methods clearly work best
when the seasonals form a fairly smooth pattern. Fourier methods have
an advantage when we are dealing with, say, 52 weeks in a year or 168
hours in a week. The manipulation of more than 50 seasonal indices is
unlikely to give better results than a Fourier model with, say, 10
parameters. If the seasonals form a very irregular pattern within the
year, then the use of the separate seasonal index for each season
becomes advisable. Where there is ample data and time for investiga-
tion, the identification and fitting of a suitable stochastic model may
lead to good forecasts. The problem of identifying an appropriate
149
model is however far from easy, e.g. see Chatfield and Prothero (1973),
Box and Jenkins (1970). Providing good forecasts may not, however, be
the only consideration. The estimates of current trends, seasonal
indexes and the deseasonalized data provided by the seasonal index
methods and Fourier methods, but not directly by the stochastic
methods, are often important supplementary information for decision
making in seasonal situations.
References
Box, G. E. P. and Jenkins, G. M. (1970). Time Series Analysis and Forecasting.
Butterworths, London.
Box, G. E. P., and Jenkins, G. M. (1973). Some comments on a paper by Chatfield
and Prothero and on a review by Kendall. J. Roy. Statist. Soc. A 135, 337—345.
Brown, R. G. (1963). Smoothing, Forecasting and Prediction of Discrete Time
Series. Prentice-Hall. Englewood Cliffs. N.J.
Brown, R. G. (1967). Decision Rules for Inventory Management. Holt Rinehart and
Winston.
Chatfield, C. and Prothero, D. L. (1973a). Box—Jenkins seasonal forecasting. J.
Roy. Statist. Soc. A. 136, 295—336.
Chatfield, C. and Prothero, D. L. (1973b). A reply to some comments by Box and
Jenkins. J. Roy. Statist. Soc. A. 136, 345—352.
Dauten, C. A. (1961). Business Cycles and Forecasting. South-Western Publishing
Co., Cincinnati, Ohio.
Davis, H. T. (1941). The Analysis of Time Series. Cowles Commission, Yale.
Durbin, J. and Murphy, M. J. (1975). Seasonal adjustment based on a mixed
additive—multiplicative model. J. Roy. Statist. Soc. A. 138, 385—410.
Duvall, R. M. (1966). Time series analysis by modified least-squares techniques.
JASA, 61, 152-165.
Grether, D. M. and Nerlove, M. (1970). Some properties of optimal seasonal
adjustment. Econometrica, 38, 682—703.
Groff, G. K. (1973). Empirical comparison of models for short range forecasting.
Management Science, 20, 22—30.
Harrison, P. J. (1965). Short-term sales forecasting. J. Roy. Statist. Soc. C, 14,
102-139.
Jorgenson, D. W. (1967). Seasonal adjustment of data for econometric analysis.
JASA, 62,137-140.
Kendall, M. G. (1973). Time Series. Griffin Co. Ltd., London.
Kendall, M. G. and Stuart, A. (1966). The Advanced Theory of Statistics, Vol. 3.
Griffin, London.
Lovell, M. C. (1963). Seasonal adjustment of economic time series and multiple
regression analysis. JASA, 58, 993—1010.
Nettheim, N. F. (1964). Fourier methods for evolving seasonals. JASA, 60,
492-502.
Reid, D. J. (1972). A comparison of forecasting techniques in economic time-series.
In M. J. Branson and others (Eds.), Forecasting in Action. Op. Res. Soc. and
Soc. for Long Range Planning, London.
Slutsky, E. (1937). The summation of random causes as the source of cyclic
processes. Econometrika, 5.
Chapter 9
Growth curves
9.1 Models for growth curves

The aim of this section is to introduce some non-linear models that are
finding increasing use in forecasting. These models are used in situations
with an element of growth (or decay) in them. Examples of such curves
are shown in Figure 9.1. The features of these curves, that cannot easily
be modelled by the use of polynomials, are the regions of very rapid
change and of maximum (or minimum) values to which some of the
curves tend. Let us start by considering a number of situations where
such curves might occur.
(a) Many items, such as consumption of electricity or numbers of
books published, seem to increase at approximately a constant
percentage rate, e.g. a 5 per cent increase each year. Thus if the
first year’s sales are 100 units, the second year would be
100 x 105/100 = 105 and the third year would be
105 x 105/100 = 110.25. The plot of this data extended is shown
in Figure 9.2 on ordinary graph paper. Figure 9.2 also shows the
line of the logarithms of the data. The series of original data curve
upwards as sales expand. The logarithms of the data follow a
straight line. The reason for this is that we could write the third
year’s sales as 100 x (105/100)2, the fourth as 100 x (105/100)3,
etc.
Taking logarithms gives log sales as
105
log 100 + 2 log
100
105
log 100 + 3 log
100
etc. Thus the log sales increase by a constant amount log 105/100
150
151
Figure 9.1 Sketches of some typical growth curves
each year. We could write the general model for such sales as
xt = arr
or
log xt = log a + t log r
where in the above example a = 100 and r = 1.05. It is often

convenient to write r = e13 and use logarithms to the base e. The
above models thus become
xt = ae^
called the exponential growth curve, and
log xt = log a. + fit

This model gives a good description of many sets of data and can
be very good at producing short-term forecasts. However, when we
152
Data
Figure 9.2 Plot of data and log data for exponential growth
consider longer term forecasts, there must always be an element of

doubt that any real series will continue to increase with greater and
greater rapidity, as is implied by the exponential curve. It is more
likely that the rate of growth will die off as time goes on and,
possibly, that some sort of saturation level will eventually be
reached.
(b) Situations where the figures reach some fairly steady maximum or
minimum value are common. The lines to which these curves tend,
shown dotted in Figure 9.1, are called asymptotes, and these three
curves are sometimes called asymptotic growth curves. It often
happens that when a new product is launched, which competes
with existing products in its field, it will expand into the field from
zero sales until it slowly ceases expansion at some certain
percentage of the total sales in the field. Thus a model for the sales
of this product during the expansion phase as a percentage of the
total market would most appropriately be one of the asymptotic
growth curves.
(c) Occasionally we have an item, such as a stylized product, that will
153
only sell for a certain period, e.g. a year in the case of fashion
clothes. The monthly sales for such an item might follow a curve,
such as the logarithmic parabola in Figure 9.1, or, alternatively, we
might try to model the total sales up to any given month by one
of the asymptotic growth curves.
The growth curves of Figure 9.1 thus provide a useful class of

models. Lewandowski (1974) discusses a large number of growth curves
of use in forecasting. Those mentioned in Figure 9.1 are the most
common. It is not proposed to carry out a detailed mathematical study
of these curves individually. In Table 9.1 the main formulae and
properties associated with these curves are given. They are all described
by models with at most three parameters. Though these parameters can
take any values, their use in prediction tends to impose conditions on
the parameters, as shown in the table. The condition shown without
brackets has been assumed when calculating the values of xt and the
slope when t approaches minus and plus infinity. Table 9.1 illustrates
the significance of some of the parameters in terms of the shape of the
curve. 'Referring to Table 9.1 and Figure 9.1, it is clear that the
Gompertz and logistic curves are very similar. The meanings of the
various parameters are intuitively the same: a gives the position of the
asymptote, y controls the scale of the graph along the time axis and (3
the positioning of the point of inflexion.
The problem of deciding from a set of data which curve is
appropriate, the identification problem, is particularly difficult in
forecasting. One’s initial data usually come from the left-hand end of
the axis and at this end the curves look pretty much the same. There
are two common ways of deciding which model is appropriate. The
simplest is to plot the data on special graph paper using axes designed
to give straight lines if the data come from the appropriate model.
A second approach to identifying the appropriate model is to make
use of what are called ‘slope characteristics’. If we calculate the slope
and express it in terms of xt and t, we obtain the expressions given
under the heading ‘slope equation’ in Table 9.1. If these equations are
rearranged so that the left-hand side takes the form given under the
heading ‘slope characteristic’, then the right-hand side has the form of a
linear trend, possibly with zero slope. For example, for the logarithmic
parabola
slope/x = (3 + 2 yt
and for the Gompertz curve
log (slope/x) = (log y(3) + 7t
Thus if we plot the various slope characteristics against time, then we

will get a straight line in the one case where the slope characteristic is

• r—i
%->
CD
a> CD a>
a, (X a,
o o o o
'l) cd a> cn a> To To
a. %
03 a, a
o
X, X o tuo o OJD tuo
co a O To o O
•K. n *«i
s X *
CO
c- X a>
11
CM
W X
■K> x + CD ^1 «
co.
co- oa i
c
o ll II II ll II
QJ *-4^ ■K* •k* -k.
a> CD a> a> a>
a ^ CU a a, a a,
o ^ cr1 j3 o o o o
co a> To To To To To
O O
II II
Cl
t _i
* 5 - ■+- CD *5^ CM
o .2 c,
X -+~i 0)
CD CD CO.
c ai ^ 3 G O ii ^ II
-X
'SC ci O O £ +
d
Cl
x
<2.S£ Z Z H rH X
Table 9.1 Growth curves
a> 8 o 8 o o o
a
8 o
*1A
t X 8 CO 8 o 3 s
ts
ca
<D
+
*§■ t t t # f
a»
a.
o
O To s ca.
H + 'Q. +
II OJ
CO S3 S
a
o
o
c<3 XJ
3 C3
o
A
A
Q°
vY
s—✓
A VA
,
o ° O ° °
V A v
°?°
V A A
O <3
i—* o ca CO o
154
155
the right one for the model. A detailed study of the use of slope
characteristics is given by Gregg, Hassell and Richardson (1964).
The method has two practical difficulties whose importance depends
very much on the smoothness of the data available. There is first the
problem of measuring the slope at different times. This can be done
approximately by taking first differences and smoothing these with a
moving average. Alternatively, one can smooth first and then take
differences or find the slope of a fitted trend line. Neither approach is
very satisfactory. The second problem comes from the fact that the
method depends on an eye comparison of different plots to see which
looks most like a straight line. One can, however, be misled in this,
since the vertical scales are all in different units. Within the forecasting
context it is best to follow a preliminary study of the data, using one of
the above two methods, by a study of the quality of forecasts that the
most appropriate models produce.
An important property of the growth curves of Table 9.1 is that by
suitable choice of transformation they can be transformed either to a
linear model or to a simple modified exponential. The nature of these
transformations is displayed in Table 9.2. As long as we are investi-
gating deterministic models, these transformations can be quite useful
in helping us to inspect the data to decide what model is appropriate;
we could also use the linear form as the basis of extrapolation. The
same is also true if there is a very small random component in the
model. If, however, there is a large random element superimposed on a
Table 9.2 Transformations of growth curves
Growth curve Linear form
Exponential log xt = log a + Pt

Simple modified exponential log (xt — 5) = log a + pt
Logarithmic parabola log xt = log a + pt + 712
Simple modified exponential form (and

linear form)
Gompertz curve log xt = log a + j3e7f

and hence log (log xt — log a) = log p + yt
Logistic curve t = i + -ei"

x a a
( 1 1\ P
and hence log I I = log — + yt
\xt aj a
Note that log here is loge, the natural logarithm.

156
growth curve, the use of transformations to linear form, followed by

estimation of parameters and extrapolation, may lead to poor forecasts.
9.2 Fitting growth curves
There is a large literature dealing with the fitting of growth curves (e.g.
Harrison and Pearce, 1972; Lewandowski, 1974). Rather than attempt
the rather daunting task of giving a survey of the subject, the aim of
this section is to indicate by examples a number of approaches that can
be adopted.
9.2.1 Transformation, fitting and re transforming
This approach is very common and is based on the transformations

shown in Table 9.2. For example, a set of data from an exponential
situation would be transformed by taking logs. A least squares or
discounted least squares line is then fitted as described in Chapter 5 and
antilogs are then taken to give the fitted exponential growth curve. The
fitting of the logged data assumes that the model being used is
loge xt = loge a + 13t + et, Var(ef) = o2
i.e.
€f
xt =
t +
Now we normally regard the form ae^ as giving the expected curve
with random deviations from it. Unfortunately, the expectation of xt in
the above model is ae^t+u /2 and not as is implicitly assumed by
the above method. Thus even if we obtained a perfect fit and forecast
xt+h by
Xf + h = ae^t+ h}
we would have a bias given by

2
E(@t+ h ) —
E(Xt+ h %t+ h ) ae P(t + h) ê<7 /2 —
1)
This bias increases with a2 and also increases as one climbs up the
growth curve. We will discuss in Chapter 16 methods to adjust for
systematic bias in forecasts. For the moment it is sufficient to draw
attention to the problem of the possible introduction of bias when
transformations are used.
9.2.2 Least squares and discounted least squares
Consider as a model for exponential growth the expression
xt = oce^t + et
157
Figure 9.3 Contours of S
The method of least squares requires us to choose estimators a and jS to

minimize
S = 2 e2 = 2(xf-cee^)2
t t
We can do this by a procedure that simply works on the assumption

that 5 is a surface in the (&,|S) plane, such as that shown by the contours
in Figure 9.3. Various pairs of values (dq ,/5j), (&2,jS2), • • • are substitu-
ted and the values Sx, S2, . . . calculated until the minimum point Smin
is found. The corresponding values (&min,jSmin) are then the least
squares estimators. Exactly the same procedure may be adopted if we
use a discounted least squares criterion.
There are two classes of method for choosing the sequence of points
(cq ), (d2,/32) in the search for the minimum. In the first class,
methods are devised for moving towards the minimum by obtaining S
at a sequence of points and using these values to guide us to where to
move to next. Such methods are often termed ‘hill climbing’ methods.
The second class of method makes specific use of the form of the
function, the growth curve. Thus the slope of the curve being fitted can
be used to choose a suitable sequence of points. Draper and Smith
(1966) discuss three such methods. The method of Marquardt (1963)
and its various developments are particularly relevant to fitting growth
curves (e.g. see Barham and Drane, 1972).
If we try to find (&min,(5min) theoretically by using calculus, we
obtain the following pair of normal equations:
2 xtoPt — a 2 e2 ^ = 0
t t
2 txte— a 2 te2 ^ = 0
t t
158
which do not have a simple solution. If, however, we can obtain an

estimate jSj of j8, then the first equation gives
cq = 2 xte^ {/2 e2 ^ x
t t
which can be evaluated directly. The second equation can be used to

give a check on the closeness of the solution to the least squares
solution and to produce a revised estimate of (3; the process can then be
repeated until satisfactory solutions are found.
In both least squares and discounted least squares it is assumed that
the residuals e,- have the same variance for all t. It will often occur that
the variablity will increase with the level of growth curve. In this case
the methods should be modified to replace the term e] by wtef, where
u)i = 1/(estimated variance of xt)
The intuitive logic of this is that if the variance is large less notice
should be taken of a large One way of obtaining such an estimated
variance is to model the variance as a function of level of the curve. For
example, one method uses
Variance is proportional to (level)k
where the level of the curve is found using a weighted moving average
of the data and k is some power determined to give the best fit (see
Harrison and Pearce, 1972).
9.2.3 A data sectioning method

An approach that can be applied to several of the growth curves to
obtain simple approximate fits is to divide the data into sections and
work out some statistics for each section. The unknown parameters are
then found from these statistics. To illustrate this approach consider
the modified exponential
xt = 5 + ae^ + et
A
The normal equations for this can be solved simply for 5 and &, but the
term is not separable. If, as before, we can get an estimate of |3, then
the normal equations can be used to provide estimates of 5 and a.
Suppose we have 3 n observations and we consider the model values
xl = 3 + ae^ + e j
x2 = § + cee2 ^ + e2
xn = 5 + aen^ + en
x j = 3 + ael(S + e {
159
where xl is jthe average of the first n observations, ey of the first n e

values and el(5 is the average of e^, e2(S, . . . , en(S. If we do the same for
the next n observations, we get
x2 - 5 + oten P el@ + e2
since the terms in etl3 all have the form e(n + *^. Similarly, for the last n
observations
x3 = 5 +ae2nêl(S + e 3
Replacing the e by their expected values of zero, we can eliminate 5 by

using
*2~ cxe1 P en P (en P — 1) (—means approximately equal to)

and
x 2 — Xj — ael(5(en<3 — 1)
so we have as an estimator
A
n(i
e = (x 3 — x2)/(x2 — x1)
This can be used with the equations above or the normal equations to
obtain estimators of 5 and ce. Gregg, Hassell and Richardson (1964) give
details of these methods and tables to simplify the calculations.
9.2.4 Other ways of producing a linear form
The form of the slope equation can suggest other ways of obtaining
linear equations. For example, in the case of the logistic curve we can
remove the eyt term by using the basic equation of the model and
obtain
— slopef = 7 — — xt
xt (x
which we can write as

zt=y — 6xt
which is an ordinary linear model. The values of zt can be found,

roughly, from the data using the methods suggested for finding slope
characteristics. We can then apply ordinary or discounted least squares
to estimate 7 and 9; a then comes from —y/6. To estimate (3 we note
from the model, putting et as its expected value of zero, that
pê~yt
So if we define
an estimate of (3 is then given by the average ut thus:
0 = 2 Ui/t
1=1
9.3 Forecasting growth curves
Much of the literature on forecasting makes a distinction between

long-term and short-term forecasting. From the statistical viewpoint
this is not a meaningful distinction, as all our forecasting is based on
model-fitting and extrapolation. Whether we extrapolate for a lead time
of 1 or of 100, the basic procedure is the same. Ideally we choose the
best model for both lead times. However, if we leave statistical
considerations aside, then it is clear that, for a lead time of 1, several
different models may well give similar forecasts and a choice of model
based on simplicity of use might well be appropriate. Conversely, for a
long-term forecast, a model that correctly follows the curvature and
possible asymptotic properties of the data will justify the extra effort
required to obtain a good fit. A consequence of these considerations is
that growth curves have mainly been used for long-term forecasting.
For such a use considerable care must be given to the consideration of
(a) the structure and (b) the stability of the situation.
(a) For the structure we need to identify not only the type of growth
curve to use but also the way the random variation is involved in
the structure. For example, an exponential growth might be of the
form
xt = ae^f + et
or
xt = ae^t+ e ^
In the former case least squares methods should be applied to the

data directly; in the latter case to loge xt. In this latter case the
model for log xt is the ordinary linear trend model of Chapter 5.
Having obtained estimates a and (3 the natural forecast will be
xt+h = aetf(f+',)
in both cases. As indicated before, the non-linear form of the

model means that the forecasts xt+h will in general be biased, the
amount of bias increasing with o]. There are no clear-cut ways of
identifying and using the ‘right’ structure. If adequate historical data
161
is available, several alternatives should be considered and the forecast

errors analysed using the methods discussed in Chapter 15.
(b) In studying the likely stability of the model a thorough study is
needed of the nature of the variable being forecast. For a growth
curve to be assumed stable over a long period, the growth shown
must be a natural result of an existing and developing structure of,
for example, market forces. For long-term forecasting the assump-
tion of stability needs to be related to the total environment in
which the model operates. Though a study is beyond the scope of
this book, it is worth noting that the types of growth curve in
common use can gill be derived from more general models of some
situations. These models describe the ways in which the variable xt
relates to various forces that act upon it. For example, suppose xt
are the weekly sales in a market with ultimate potential weekly
sales of S. A possible model would assume that the rate at which xt
will increase will depend on the number of potential customers,
who do not yet buy the product. Thus xt will increase most rapidly
at first but slow down as S is approached. The model is thus
Rate of change of xt = fi(S — xt)
where p is a constant of proportionality. A mathematical analysis

of this shows that xt must follow the growth curve.
xt =S( l — e-fit)
which is the special case of the simple modified exponential curve
illustrated in Figure 9.1.
References
Barham, R. H., and Drane, W. (1972). An algorithm for least squares estimation of
non-linear parameters when some of the parameters are linear. Technometrics,
14, 757-766.
Draper, N. R., and Smith, H. (1966). Applied Regression Analysis. John Wiley and
Sons, New York.
Gregg, J. V., Hassell, C. H., and Richardson, J. T. (1964). Mathematical Trend
Curves; An Aid to Forecasting. ICI Monograph. Oliver and Boyd. Edinburgh.
Harrison, P. J., and Pearce, S. F. (1972). The use of trend curves as an aid to market
forecasting. Ind. Mark. Manage., 2,149—170.
Lewandowski, R. (1974). Prognose und Informations Systeme. W. de Gruyter,
Berlin.
Marquardt, D. W. (1963). An algorithm for least squares estimation of non-linear
parameters. J. Soc. Ind. Appl. in Math., 11,431.
Chapter 10
Probabilistic models
10.1 Introduction
In previous chapters interest has centred on forecasting some variable
xt. In a number of situations the item of central interest is an event or a
number of possible events. The model used then concentrates on the
probabilities of these events occurring. The forecasting problem in such
situations is that of estimating the future values of these probabilities.
For example, the prices of certain foodstuffs tend to be either stable,
with the basic price remaining constant from day to day, or unstable,
with price changes taking place. In obtaining price forecasts we are
faced with forecasting the probabilities of the two alternative events,
‘prices stable’ or ‘prices unstable’, at a future time. Another situation
which faces us more directly with probability considerations is where
we are forecasting a discrete variable that may only take on a limited
range of values. For example one might require to forecast, in a
changing situation, the number st of vehicles out of a small fleet of five
that would be in use at a future time. Such a problem requires a
‘forecast’ of the future probabilities of demands occurring for use of a
vehicle. The aim of this chapter is to discuss briefly some approaches to
the forecasting of probabilities.
10.2 Forecasting probabilities

We now consider the problem of how to estimate probabilities in
situations where there may be slow changes in these probabilities. We
are dealing with a situation, analagous to that discussed in Chapter 4,
where there are simple and well-known methods for finding estimates
of probabilities where these are globally constant, but where we must
now regard the model as only locally constant.
In previous chapters we have frequently made use of ‘discounted
least squares’ to fit a model locally for forecasting use. In this approach
162
163
the usual least squares criterion of fit
X (residual)?
i= 1
was replaced by
t- 1
X icr(residual)?_ r
r= 0
where the residual was the error between the observed value and the
fitted value and the weights wr were chosen to give more emphasis to
recent data than past data. In fitting probability models the method of
least squares is often replaced by a method called the method of
maximum likelihood. The basis of this method is first to write down
the probability of the events that were observed. This probability is a
function of the model parameters and is called the likelihood. Next we
choose the model parameters so that the events observed had a higher
probability of occurrence than events that were not observed. Thus if
the probability of observing the discrete variable st at time t is Pt(st;0),
where 6 is a parameter, the likelihood, probability, of observing data
si . . . Sf is
<?(d)=P1(s1;6)P2(s2 ;6) . . . Pt(st;d)
We now choose the parameter 6 to maximize this probability. Another

way of expressing this is to take logarithms of both sides giving
L(6) = logir(0)= i log

1= 1
The value of 6 which maximizes this also maximizes Thus the

estimator 6 can be chosen by maximizing £?{d) or L(0), whichever is
most convenient.
A natural way to modify this global method of estimation, to deal
with local changes in parameter values, is to introduce a discounting
factor wr, as above. The discounted versions of £f(d) and L(0), denoted
by (6) and Lw (6) and called the discounted likelihood and discounted
log likelihood, are
&w(6) ={Pr(sI;e)}lv«{P,_ , (sf_ ,; 0)}w ■.. .{P, («,; 0)}Wf- 1
and
Lw{9)= 2 wr\ogP,_r(st_r;d)
r= 0
It will be seen that the discounting is applied to the term log

Pt_ r(st- rm90) in exactly the same way as it is applied to (residual)?_ r in
discounted least squares.
164
Let us illustrate this method by considering the food price example

of section 10.1. Let p be the probability of the price being stable (S) at
any time and, hence, 1 — p be the probability of it being unstable (U).
Thus Pt _ r(St_ r;6). in the discounted likelihood, is just p when S occurs
and 1 — p when U occurs. If we use our simple exponential discounting
weights, wr = </, the terms in the discounted log likelihood are as
shown for a particular set of data in Table 10.1(a). We can obtain a
local estimate of p, p, by finding the value of p that makes the quantity
Table 10.1 Local estimation of a probability
(a) Theory
Stable S or
Time unstable U Probability Contribution to Lw(p)
t S P logp
t—1 S P a log p
t-2 s P a
2
logp
t-3 u 1 ~P a 3
log(l -p)
t—4 s P a 4
logp
t-5 u 1 -P a5 log( 1 -p)
t-6 u 1 -P a6 log(l -p)
t—1 s P a7 logp
t-8 u 1 ~P a8 log( 1 -p)
t-9 u 1 -P a9 log(l -p)
Total Lw{p)
pt maximizes Lw(p)
(b) Calculation (not using a recurrence form)
G = 1 a = 0.9
Time Sort/ ar arbt-r ar arbt_r ar arbt-r
t S 1 1 1 1 1 1.000 1.000
t—1 S 1 a a 1 1 0.900 0.900
t—2 S 1 a2 a2 1 1 0.810 0.810
t-3 u 0 a3 0 1 0 0.729 0.000
f-4 s 1 G
4
G
4
1 1 0.656 0.656
f- 5 u 0 G
5
0 1 0 0.590 0.000
^— 6 u 0 G
6
0 1 0 0.531 0.000
f-7 s 1 G
7
G
7
1 1 0.478 0.478
t—S u 0 G
8
0 1 0 0.430 0.000
t-9 u 0 G
9
0 1 0 0.387 0.000
Totals A D 10 5 6.511 3.844

3.844
= 0.590
Pt = D/A Pt 5/10
6.511
= 0.5
165
Lw(p) as large as possible. This looks a little complex, but we can

simplify the formulation of the problem by introducing a quantity
called an indicator variable 5?_r. This is set at one or zero according to
whether S or U occurs at time t—r, as is seen in Table 10.1(b).
The contributions to Lw (p) can now be written as
ar[8 t — r logp + (1 — r) log(l —P)]

Maximizing Lw (p) with respect to p can now be shown to lead to
Thus pt is the discounted average of the indicator variables. Table

10.1(b) shows the fairly simple layout required to calculate pt at a
fixed point in time. As usual D and A can be simply updated as time
goes by. If a = 1, then pt is the ordinary average of the indicator
variables, i.e. the proportion of occurrences of the stable state. In Table
10.1(b) for a = 1, we have p = 5/10 - 0.5, as for half the weeks the state
is stable. If a - 0.9, this becomes 0.59, indicating the preponderance of
stable states in recent weeks. The above expression for p, though
derived for the case of two possible events, is in fact quite general. We
obtain the local estimate of the probability of any event, of a number
of possible events, from this formula by using 5 = 1 when the event of
interest occurs and 5 = 0 when any of the other events occur.
When t is large we can rewrite the expression for p approximately as
Pt = (1 — a) 2 ar8t_r
r= 0
or, in recurrence form,
+
Pt = (1 — a)§r apt- 1
This form was suggested by Brown (1963) on rather different grounds.

Experience with using these methods suggests that in most situations
it is advisable to take a value of a fairly close to one; otherwise the
estimates p tend to fluctuate too much at random.
A second example of the use of this approach is given in Figure 10.1.
This shows the use of the numerator, 2ar5?_r, of p for the occurrence
of the discrete variables S = 1, 2, 3 or 4 over a period of time. For a = 1
this is just the frequency of occurrence of the particular values of S,
which are normally plotted as a bar chart. For the value a = 0.99 used it
will be seen that the plots of the discounted frequencies show up a
changing pattern in the frequency distribution. A variation on this plot
can be used to observe any changes in the proportion of large errors
occurring in an ordinary forecasting system.
166
\
a = 0 99
After 30
observations
After 45
observations
After 60
observations
12 3 4
Figure 10.1 Example of a discounted frequency

bar chart for a changing situation
10.3 State models
In this section we will briefly look at some forecasting situations where

the interest is in the probability of the situation changing from one
state to another. For example, in the price forecasting problem
described in the last section it is found that in practice the probability
of being in the stable state S next week depends on whether we are in
the stable state S or the unstable state U this week. We are thus
interested in the occurrence of transitions from one state to another.
Table 10.2(a) shows the number of transitions of the four possible
types for a particular set of vegetable prices. It will be seen that if we
are in state S this week we are twice as likely to be in state S next week
as in state U. In probability terms the ‘transition’ probability of going
from state S to state S is about 2/3. Table 10.2(b) gives the notation for
the possible transition probabilities and Figure 10.2 gives a diagramatic
representation of such a situation.
A situation like this which takes the form of a ‘chain of events’ with
a limited number of possible states is called a Markov chain. The theory
and properties of Markov chains have been used extensively in
forecasting, particularly within the market research area of application.
167
Table 10.2 State models
(a) Number of transitions of state
State in next week

S U
State in S 35 18
first week u 11 36
(b) Notation for probabilities of

transition
State in next week

S U
State in s Pi 1 “Pi
first week u 1 ~P2 P2
(c) Global estimates of the

transition probabilities from (a)
State in next week

S U
State in s 0.66 0.34

first week u 0.23 0.77
Figure 10.2 Transition diagram
One problem here lies in the estimating of the transition probabilities.

It can be done globally as above by simply counting observed
transitions. Thus of the 53 transitions from state S, 35 are to S, giving
an estimate px =0.66. Table 10.2(c) gives the four estimates. In the
price forecasting situation these transition probabilities vary during the
\
168
various seasons of the year and so px and p2 are best estimated locally
using the discounted method of the previous section. Figure 10.3 shows
the local estimate of px and p2 over a period of time for some price
data. Notice that we only get information about px when a transition is
made from state S to S or U\ thus only one of the curves varies at any
one time. To illustrate how to use such transition probabilities, let us
suppose that we know that we are in the stable state this week. We will
be in state S next week with estimated probability px and state U with
probability 1 — px. The estimated probability of being in state S two
weeks hence is that of going through the sequence of states S,S,S or
through S,U,S9 which is px x px + (1 ~px) x (1 — p2).
Before leaving Markov chains we consider a further forecasting
problem that is similar in structure to the above. By way of example,
Table 10.3 shows some information about a steady market in which
four brands are competing. Part (a) gives the current market share of
the four products. Part (b) gives the transition probabilities of product
buyers between the brands. Thus buyers of brands A and B are
completely loyal, but a person who buys brands D in year t has only a
probability of 0.48 of buying it in year t + 1. Part (c) shows how this
information can be used; to forecast the brand shares in year t + 1. We
assume that because brand A held 50 per cent of the market in year t
and had totally loyal customers it will hold an initial 50 per cent of the
market. To this we must add the fraction 0.02 of the 10 per cent of the
market that used to buy C, but which switches to A, and the fraction
0.22 of D brand’s 10 per cent of the market, giving a total of 52.4 per
cent. A basic problem here is that of estimating the probabilities of the
transitions. One way of doing this, discussed by Chorofas (1965), is to
carry out the reverse of the above calculations over a period of years for
which market shares are known in successive years.
169
Table 10.3 A market share example

(a) Market share year t (%)
Product
A B C D
50 30 10 10
(b) Transition probabilities
To
From A B c D Total
A 1.0 0 0 0 1.0
B 0 1.0 0 0 1.0
C 0.02 0.03 0.95 0 1.0
D 0.22 0.30 0 0.48 1.0
(c) Contributions to market share year t + 1
Year t
shares Contributions to
' (%) A B C D
A 50 50 0 0 0
B 30 0 30 0 0
C 10 0.20 0.30 9.50 0
D 10 2.20 3.00 0 4.8
00
Total 52.4 33.3 9.5

year t+ 1
market shares
(%>
In the price forecasting example one state corresponded to the

stable, constant price situation, S, which is modelled by
xt = M
The other corresponded to the unstable price situation, U which in fact
was modelled by
xt = xt_ j + et
This example introduces the concept of using different possible states

to describe situations in which changes of structure occur. A major
application of this concept is to models such as that for the wandering
mean introduced in section 7.6, i.e.
+
xt = Ht
Vt = Mf- I +71
170 \
In the stable state S, the et and yt values come from distributions with
zero mean and variances cl and oy, respectively. If a very large value of
et occurs, then this shows as a transient in xt. If a large value of yt
occurs, pt will jump to a new level. Future means will start to wander
from this new level; thus a major step change will have occurred in the
mean level. The occurrence of such large values ofêr and yt can be
modelled by assuming the existence of two further states, Ue and Uy.
In one, Ue, the variance of et is substantially larger than o\, thus
allowing the possibility of the transient, large et. Similarly the other,
Uy, gives a large variance to yt, thus modelling the step-change. Clearly
the system will only rarely be in states Ue and Uy. The probabilities of
the transitions S Ue and S Uy will thus be very small, and those for
Ue+S and Uy -* S correspondingly high. Harrison and Stevens,
using Bayesian methods, investigate a particular model of this type,
which includes also a state for trend changes, U5. In their examples the
variances involved in Ue and Uy are of the order of 100 times larger
than ol and oy, thus the occurrence of states Ue or Uy represent
significant events in the progress of the data.
The value of using the state approach here is not that the forecasts
make some allowance for future changes in state. This allowance must
inevitably be small since the changes are relatively rare. The real value is
that the model, as it is fitted to the data, detects past changes in state.
This automatically improves the quality of the forecasts. Thus if a
transient is detected, it will be treated as a transient and future
forecasts will largely ignore it. Many other forecasting approaches will
treat the transient as a respectable observation and it will lead to biased
forecasts. Similarly, if a step-change is detected, the new forecasts made
will be based on the levels of pt after the step and will tend to ignore
data from before it. Other forecasting approaches will often take some
time for the estimated level to move over to the region of the new level.
The model incorporating the possibilities of transient and step-change
states thus deals more effectively with data in which such features can
occur. For technical details the reader is referred to the reference
already given.
10.4 Forecasting criteria
In this chapter we have been examining the forecasting of one or more

probabilities pl,...,pk using some forecasts p1,...,pk. If our
forecasts are sensible they will satisfy 0 < p, < 1 and 2f= j p{ = 1. It will
usually occur that when we have obtained our observations we will only
know whether certain things have happened or not; we will not know
Pi. The consequence of these facts is that our usual mean square error
criteria of a good forecast has no direct application. It is thus necessary
to develop some new criteria upon which to judge forecasts. Let us start
171
by considering the type of situation where pt represents a proportion

that will in fact be observed. In this case we can compare the forecasts
Pi with the observed pt. To derive such a comparison we use a method,
given by Theil (1965), based on information theory. Suppose the event
Et has probability pt and the statement is made ‘E{ has happened’. If
Pi = 1 this statement does not really tell you much: we already knew it
was almost certain to happen. If, however, pt is small, the statement is
much more informative. The amount of information It in the statement
is clearly an increasing function of 1 /pt and in fact we use
Ii = log 1 Ip,
as a measure of information. If we have k possibilities, the expectation
of the amount of information to be obtained from our observations is
k
I = 2 Pi log 11pi
i= 1
k
= — 2 Pi log pi
i= 1
Similarly, the expectation of the information in our forecasts pt is
A *
/ = - 2 Pi log pi
i= l
If we had perfect forecasts these would be identical. The difference in

information content of p and p is thus a measure of forecast accuracy.
This is given by
- A * Pi
D = / — / = 2 Pi log —
■-1 Pi
The smaller this quantity the better the forecasts, perfect forecasts
giving D - 0. Similarly, two different forecasts pt and pt can be
compared using D — D, i.e.
f 7 * ] Pi
I —T= 2 Pi logr
i= 1 Pi
If this is positive, p will be judged the better forecast and conversely if

it is negative. If we have no knowledge of but only the information
that event Ej had occurred, we could assign py = 1 and all other pt = 0.
This would be equivalent to judging that method as best which assigned
the highest probability to Ej. If we have several sets of observations, we
would simply let pt be the observed relative frequency of occurrence of
Et (i = 1, . . . , k).
Suppose that we have a sequence of observations at time 1, 2, . . . , f,
with respective probabilities pl5 p2, . . . ,pt. The information content
would be
t
— 2 log Pj
r= 1
If we were concerned with choosing parameters in p for forecasting

purposes, as in section 10.2, we would wish to emphasize the most
recent information. We therefore define the discounted information
with weights wr to be
t— I
— 2 u)r logpf_r
r= 0
It has already been noted that if we know the future and could assign
Pi = 1 to the appropriate future event Eh then the information content
of this perfect forecast would be zero. Thus we are aiming to produce
forecasts with information at a minimum. This suggests that we choose
the parameters in the formulae for the probabilities to minimize the
discounted information. Referring back to section 10.2, this is just the
same as maximizing the log discounted likelihood which is
t- I
2 wr\ogpt_r
r= 0
This was the method used to obtain our forecasts in that section. Thus
our brief discussion of information has given further support for the
methods of obtaining probability forecasts in section 10.2.
10.5 Probability forecasting and decision making

The above discussion assumed that our interest centred in the
probabilities themselves and that we had no further information about
the practical situation. The usual object of calculating forecast
probabilities pt is to enable a decision d to be made. Let dt be the
decision to be made if one was certain that the event Et would occur.
Let Ctj be the cost involved if decision dt is made and event Ej actually
occurs. The expected cost of decision dt is
2 Cup/
/= 1
Pj being the true probability of Ej occurring. As we do not know this,

we must use Pj as our best forecast of this future probability. The
estimated expected cost of dh Cj, is then given by
173
Table 10.4 Calculating decision costs (cost table)
Estimated
expected
Events occurring costs
Forecast events E, ... E-j
J-! . . . Ek Decision C
E, Cn • • •
r..
W • • • Ci* d\ Ci
Ei Cn • • • C-
V/^y • • •
E'ik dt Ci
Ek Cki • • • E>kj • • • Ek k dk Ck
Forecast probabilities Pl • • • Pj •• • Pk
Using a layout such as that shown in Table 10.4, the value of Ct can
be calculated for the k possible decisions. It is then often a reasonable
procedure to choose the decision dh corresponding to the smallest
estimated cost Cm in. If we wish to compare two different sets of
forecasts, p and p, one could compare the corresponding sets of
estimated expected costs C and C One would certainly be interested to
see whether they both lead to the same decision. Suppose, for example,
there are only two possible decisions, dx and d2 ; then it is clear that, on
the basis of the table, when p1 is less than some value, p0 which
depends on the costs, we will reach decision d2 and conversely when
pl > p0. Thus any forecast will lead to the same decision provided it
lies on one particular side of p0. Consider another example of the use
of these expected cost tables. Suppose that the costs associated with
decision dt when E) occurs are constant, C say, for all i,j(i =£ j). If, also,
the costs associated with the correct decision (i = j) are all at some
constant value less than C, then the best decision is the one that
corresponds to the most probable event as indicated by p. This if p3 is
the largest probability, d3 is the decision made.
References
Brown, R. G. (1963). Smoothing., Forecasting and Prediction of Discrete Time
Series. Prentice-Hall. Englewood Cliffs, New Jersey.
Chorofas, D. N. (1965), Systems and Simulation. Academic Press, New York.
Harrison, P. J. and Stevens, C. F. (1971). A Bayesian approach to short-term
forecasting. Op. Res. Quarterly, 22, No. 4, 341—362.
Theil, H. (1965). Applied Economic Forecasting. North-Holland Publishing Co.,
Amsterdam.
Chapter 11
Multivariate models
11.1 Introduction
Most people involved in forecasting have to forecast more than one

variable. They thus in practice have to carry out multivariate (many
variable) forecasting. For example, a sales department will usually be
involved in forecasting the sales of many different products and also of
the various versions of certain products that are made in different sizes,
colours, etc. Usually each variable is forecast independently. There are
many situations where this is done because it is expedient rather than
natural. It is far simpler to treat each series separately than to get
involved in the complexities of real multivariate forecasting. Certainly
multivariate forecasting is a fairly sophisticated subject and most of it is
beyond the scope of this book. For an advanced study see Hannan
(1970) and Robinson (1967). In this chapter we will aim at introducing
some of the ideas of the subject without getting too involved in
mathematics. However, there is a real need to recognize the possible
value of multivariate forecasting. Consider, for example, the following
situation.
In certain markets price is controlled by the market rather than the
seller. In such situations price and amount sold both vary and require
forecasting. Clearly, each affects the other and to try and forecast each,
ignoring the other, would lead to poor forecasts. To obtain reasonable
forecasts we must allow for the past behaviour and interrelationships of
the prices and the sales.
When discussing univariate forecasting, we made frequent use of the
mean square error as a forecasting criteria. When we have several
variables, we will have a mean square error matrix. This is an array of
terms of the mean square error type. Thus, if el9 . . . , en are the
forecast errors in the n variables being forecast, the mean square error
matrix is
174
175
\E{e\) E(exe2) . . . E(elen)

E{e2ei ) E(e\) E(e2en)
• • • • • • • • • • • •
\E(e„e,) E(e^ ^2 ) • • • E(el)
As this consists of many elements, we cannot regard it as a single

criteria. We can imagine forecasting methods that will make some
elements satisfactorily small at the cost of having others large. It is
clearly very useful, however, to have a single criteria, and this is
provided by the sum of the diagonal elements of the matrix:
I = E(el)+E(e22) + ...+E(e2n)
This is called the trace of the matrix. If we imagine an n dimensional

space in which the point X = (xj, . . . , xn) is the observed value of our
n variables and F = (£ x, . . . , xn) is the forecast value, then (by the
theorem of Pythagoras in n dimensions) / is the mean square distance
between F and X. Seeking to make this small is thus a reasonable
approach to constructing a single criteria for multivariate forecasting.
11.2 Generalizing exponential smoothing*
So that we do not get too involved with mathematics, let us consider a

bivariate forecasting situation with pairs of observations (acj, yx), . . . ,
(xt, yt). In ordinary exponential smoothing we discussed the use of the
error correction form (section 4.2) for which
xt = xt_ ! + bet
where et is the latest error, xt — x t_ j, and b = 1 — a is a forecasting

parameter. The natural generalization of this is
xt =xt„1 + bixex>t + bl2ey>t

= + e + e
yt y t— 1 2 1 x, t 2 2 y, t
where xt and yt depend on the errors, ex t and ey t, of the last forecasts

of both xt and yt. The constants bll9 bl2, b2l and b22 are the
forecasting parameters. If this is written using matrix notation, its
analogy with the single variable situation is clear. Thus
6iibi2
+
^21^22/
)h r
\ey,t
This method is the method of multivariate exponential smoothing. It

\
176
can be supported on similar grounds to ordinary exponential smooth-

ing. Thus
(a) It corresponds to an exponentially weighted average of the

observations with weights tending to zero in the far past (provided
the b values satisfy certain conditions).
(b) Multivariate exponential smoothing is optimum, with appropriately
chosen b values, for a set of models analogous to those for which
ordinary exponential smoothing is optimum. For example,
exponential smoothing is optimum for the model defined by
xt = Ht + et
M* =Mr-i + Vt
If xt, jit, et and rjt are now regarded as vectors, this describes a
multivariate process for which multivariate exponential smoothing
gives the optimum mean square error predictor, see Jones (1966)
and Kalman and Bucy (1961). The parameters b in this situation
depend on the variance—covariance properties of the e and 77
sequences. Jones gives a fairly direct way in choosing appropriate
values of these parameters.
11.3 Multivariate stochastic models
In Chapter 7 we considered how to forecast in situations modelled by a

number of different types of stochastic process. All these processes can
be generalized to cover analogous multivariate situations. The aim here
is simply to indicate how these processes can be generalized and how
the forecasts can be obtained for such multivariate situations. Consider,
first, the generalization of a simple autoregressive process
xt = 0 ixf- 1 + et
Let r\t be the random variable for the y process and suppose that the
present xt and yt depend on the past values of both the x and y
sequences. Then we have
+ + e
Xt=(j)\\Xt-\ 0i2lh-l t
+
3T =
02 1 Xt- 1 +
0223T- 1 Vt
as a bivariate first-order autoregressive process. The generalization to

higher orders simply involves introducing terms in xt_ 2» yt- 2 > etc.
We now turn to the moving average process
xt = et — det_ 1
and note that the latest e and 77 in the autoregressive model only
177
influence their respective x and y values. The natural generalization for

the moving average process is thus
,
= e
X
t t $\ let- 1 0 12Vt- 1
yt =Vr—d2iet-i -d22rit_
In generalizing both the above models we can increase the number of x
or e, or y or 77, terms in the expression for both xt and yt. There are
thus four numbers required to define the order of the process. If we
introduce vector and matrix notation putting
e*
yt Vt
^1 1 0 1 2\ ^ _ / 01 1 01 2
e
^2 1^2 2' ' 02 1 02 2
Then the above models become
x, = <t>xt_ 1 + et
and
xt = et- 0e,_ 1
and the multivariate forms are directly analogous to the univariate

forms. A consequence of this is that much of what was said about the
univariate models in Chapter 7 carries over to the multivariate models.
In particular, the optimum forecast, in the sense of section 11.1, is the
conditional expectation of the future multivariate observation, given all
past and present observations. Thus for the autoregressive model the
predictor is xt+1 =(f>xt, and for the moving average model it has the
form — 0ef. As before, in both cases the one-step-ahead prediction error
et = xt+ ! — xt+ ! is equal to et. Clearly, one can also extend these ideas
to mixed autoregressive moving average models and to models in which
some forms of trend, etc., can be differenced out. As a quick example,
we note that if the model takes the form:
x
? = x
/— 1 + et 0ef- 1
all the calculations given for the univariate model at the end of
section 7.5 follow through. This gives as optimum predictor
xf+i =X?+ (I —0)ef
where I is a matrix of ones, which is exactly the form of multivariate

exponential smoothing given in the previous section.
\
178
References
Hannan, E. J. (1970). Multiple Time Series. John Wiley and Sons, New York.
Jones, R. H. (1966). Exponential smoothing for multivariate time series. J. Roy.
Statist. Soc. B, 28, 241—251.
Kalman, R. E., and Bucy, R. S. (1961). New results in linear filtering and prediction
theory. J. of Basic Eng., 83, D 95—108.
Robinson, E. (1967). Multichannel Time Series Analysis with Digital Computer
Programs. Holden-Day Inc., San Francisco.
Chapter 12
Forecasting methods and models
12.1 Forecasting methods

In previous chapters we have examined a number of different models
and discussed various forecasting methods that might be appropriate
when these models describe, either globally or locally, a particular
situation. Our classification has been based on models rather than
methods. Table 12.1 gives a classification of some of the main methods
that we have used to estimate the model parameters as the basis for
extrapolating into the future. These fall into two main classes:
(a) those based on classical statistical methods of estimation, such as

least squares or on simple modifications of these, such as those
involving the use of the exponential discounting factor.
(b) those developed for particular situations, such as non-linear
situations, in a much more intuitive and ad hoc fashion. Of these
methods those which produce recurrence relations for updating the
estimates of the parameters form an important special class.
Table 12.1 gives a summary of the advantages and disadvantages of

these various methods.
12.2 Model building and identification
The approach of model building has been the underlying feature of the
methods discussed in previous chapters. Though we devoted Chapter 2
to a discussion of models, there are a number of implications from the
topics studied in later chapters that need comment. In any real
situation it is unlikely that we can only talk about the model for that
situation. We are usually in a position to reasonably use a number of
different models. The truth in a situation is invariabliy too complex for
anyone to grasp, assuming even that it exists. Different models will pick
up different features of ‘the truth’. We may model our data by a trend
179
K*")
>> 03
CUD
x £
02
£ G Q; 02
-S^ G
03
> £ _£ ’o
X
G C/2 CUD 02
03
G o x: C/2 G
03 • r^ £
C/2 03
x:
a>"2
02 02 JP
o 4-3 o3
,0 o3 C/2 G C/2 JP
,02
c/2 JP 03 o3
XH 02 02 Jp c3
O 02
'£ 5
in
>>’5
02
02
a o
•G
4-3
c
G
02
JP
02 4P
OH JP
O
CUD CT o
£ C/2 HP PH X o3
o 02 4J
c o o
O a/
o
Jp ..
03
JP 02 O
02 02 r—4
02
£
4P> C/2 03 G £ G
4P £ C/2 X5 G *o O C/2
G OH
X3 02 £
C/2 O ^ G x: G 02
§ § CUD 02 O
n « 02 o3
C/2
? § 02 4P>
02
JP g 5
G
C/2
O 4X>
w
_ ^4—I
02 O
02 4—3
PP
02
Jp
G
.X
02 33 G JP
o3 o
OH G W 02 02 G O
Cuo 02 JP
03 o) o w 02 p
a* ~
-1-3 T3 y 3 cr -Q 0 03 02 Jp Jp C/2
«2 Q OH X JP 02 02 02 G
G >3 X -S o
4P 02 02 — 02
03 pp O 0 JP
—i P c/2 C/2
X
> ttJD S
03 J02 o3 ^ 02 *p oe Jp 44>
X OH .5 o £ £ 02
(U c3 3 15
03 O 03 2 4-3> 02 _2 a 03 o3 CO O
C/2 ’ 1’> G
£ £ O 02
S *2 cO XC X2 02 C/2
02 pH C3 C3 JP J_I
5 -IP
XSl T3
03 C-J
c/2 <1 X o H o X5 £3 OH O- OH 03
c/2 o
C/2 »- PP 4-3
C/2 G 1/2 C/2 T3
c3 03 'g £ C/2
G
_r 02 £ 3 •§ 3 02 03 § 03 G
4-3 >»
*4 X — o nn iS -G 02 4-3 4—3 03 O C/2 X!
>P C/2 C/2 p • f-H
G r 4-^ 03
O Jp Jp c3 02 ’£
tf -S 13 £ o O-i 02 c3
2 8 > o C/2 . „ 02 4
Forecasting methods
-S-3 ri
GH 02 _ .5-1 02 X! s
£ 2 02 • ^ X
G 02 02 44-1 O 02
C/2 C/2 G G
02
S J; -c <£ O 03 5P G G 02 X 4-3
02 o o
8 £ 2 "§ G Jp G G
o £ C/2
C/2 a G
c«-. £ £ ^ £ £3 G
•— O Pp 02 03 02 02 02 o3 a O
03 03 02 02
02 c/2 5 S > c/2
. G X3 £ QH
'a OH X3 JP
'51 G 03 02 03
3 •£ CUO • — 4-3 OH
Jp X
03 tj o “ 4-3 .5 XJ c3
i—i r* 02* G Ô "02 0 1/3 & a X
- o x3 02 G G
Q 4-^>
’§ a x ; x ^ 1 o n ^
02 G
02 03
02 - © 4-3
X C/2 -G ©
C/2
0/
« 2 _>
>, a ^ X 0)
I ° 02
p
£ £ -
S c 3 £
C3
02
43
c«
si O G
-H> C3
PH
O 02
4-2 X!
Jp Q; 4-2
. H— CS
CUO c C/2 4—3 —< 4P> 03
-a g X3 02 X G £ g
cd
o 02 .
n a C£ C/2 G G 02 U
02 Pf
Table 12.1
-P c/2 02 CT
1
£ 02 02 cr
CUD G 02 O C/2 G o C £ „
G G 03 >».£ £ _Q OiX) G V2 02 O
s«
CUD -5 JP W3
03 O 43
> a .x OH 03 03 O G -G x: -P? 03 G
SP 03 33 xs c/2 G p3
£ G +? 3 ^ 02 02 X _Q
X O c3 X S
.£ ■£ QJ C7^ O
< O a CO w
cr
cO
c/2 -G
HG a S ^ G OP
G Q § Q go —P C/2 JP
C/2
£ c/2 <>2
a ° ■“
ago
CM CM CM CM H CO CO
>■ cu
03 0> ^3
X 1—I T—I CM CM CM rH CO CM rH CM CM CM CM CM CM CM
JP -fi • • • • •
w a o 16 L'~ 03 lO CD I> 00 (J) LO LO CO 03 OD X LO 00
JP
Jp Jp o-i o3
C/2
o
44-1
„ 02 03 03 02
.55 4-3 02 02 X3 "02 G
03 X G 02 X5 _
O X "£H »—H * ^"4 > JP
O c3
• r"H C/2
X O a, 02 ' JP O G
5-1 02 JP G '02
O JS o X 02 O
£ ^
Jp o o ^=2 X5 O £ S03 CUD
© X
o
G
a C/2 OH JP
pp G Ou £ 43
03 £ -2
£ -S C/2 X3
o3 .£
£ ^ C/2 £ O
o3 O o
^ o co 02
£ o 43
C/2 -O X 02
QH X -G G
CL •■—> 8 9^ - 02 a 13 x 02 Q2
>>■& O O O £ o £ 2 02 > .G
4-3 G 02 O O G
H £ £ O cn £ J -s £
feP
-Q
c/2
CUD 02 °5 a J 3
C/2
02 CO
Jp x g> 02
o3 02 p 02
G 4-3 HZ G
02
x cr
C/2
C
£H _> 02
Jp
o 4-3) 4-3 JP
x: O w G
4-^ C/2 02 ^
03 ’£ 02
a; .52 ^ 4-3 02
02
G
4-] Q ^ IP on
180
181
model, by a regression model on some economic variable or by a model

involving the structure of the market. These models may overlap in
some aspects; they may contradict each other in other aspects. We will
certainly compare these models. We may find that one gives better
forecasts over the last few years, but we would be unwise to discard the
rest unless they clearly do not fit the situation at all. Each model gives
its own insights into the total situation. Often the long-term value of
constructing a model lies not so much in its forecasting ability, which
may be poor and lead to its rejection, but in the insight into the data
and the structure of the situation developed during the model-building
exercise.
In Chapter 1 we classified forecasting methods roughly into three
groups: (a) intuitive methods, (b) causal methods and (c) extrapolative
methods. Though we have concentrated on methods in the third class,
it should be noted that both of the other methods make use of model
building. Clearly, the causal methods are essentially based on construct-
ing a model that shows the causal structure of the situation. Most
regression models and the more elaborate econometric models come
under this category.
The problem of identification is that of identifying the model to use
for a set of data. This is an aspect of statistics that lacks a formal
theory. One proceeds in terms of graphical methods and of calculating
appropriate statistics. In terms of practical forecasting the situation gets
even more imprecise since for reasons of cost, availability of computer
packages, ease of application, etc., one may finally opt for a model that
is not the best possible one for describing past data. The process is
rarely one of finding the best model; the chances are that we could
always go on improving our models provided that we had nothing else
to do. We usually start with a limited class of models and choose the
best fitting of these. The models discussed in previous chapters give a
choice of those most commonly found to be useful in forecasting.
Table 12.2 lists a number of these models with an indication of what to
look for in seeking to identify them.
The identification procedures described in Table 12.2 are mostly
fairly straightfoward. The one exception is the identification of
stochastic models. Several computer programs are now available that
produce the relevant information for identification, but the interpreta-
tion of this information, to identify the appropriate model, still
requires considerable skill.
The approach to identification in Table 12.2 is that of looking at
appropriate plots and statistics. Another approach, though one which
naturally combines with this, is to start with an assumed model and to
examine its appropriateness. There are two ways of doing this. One is to
start with a very sophisticated model that contains within it a number
of simpler models. For example, a model with a polynomial trend and
\
182
Table 12.2 Identification methods
Model Method
(a) Constant mean Plot xt against t.

xt = id + et Estimate p by x and plot on graph.
Calculate residuals et = xt — x.
Check these for: random scatter above and below zero,
lack of independence (calculate
autocorrelations),
lack of constant spread.
(See also methods of sections 12.3 and 15.2.)
(b) Linear trend Plot xt against t.
xt = a + fit + et Fit a line by least squares and add to graph.
Calculate residuals et = xt — a — fit.
Check these as above.
Moving average of residuals + plot of residuals will indicate
deviation from linearity.
(c) Exponential Plot log xt against t or xt on logarithmic graph paper; this
growth should give result as above.
xt = a ePf + et Fit by methods of Chapter 9. Residuals may not have
constant spread.
(d) Seasonal models Plots as in Chapter 8; see also plots of residuals and forecast
errors as in section 15.2.2(e) and (i).
(e) Stochastic models Calculate and plot using at least 50 observations.
Partial-autocorrelation0 + Autocorrelation -> Model identified
Zero after pth lag + Decaying values -> Autoregressive
process of order p
Decaying values + zero after qth lag -> Moving average
process of order q
Decaying values + Decaying values -» Mixed AR—MA
process
°For definitions see Appendix A.
several seasonal terms or a high-order autoregressive^moving average

model might be appropriate. These models would be fitted to the data,
then any terms in the model that made little contribution to the quality
of it, i.e. whose coefficients were small, would then be neglected. This
procedure is often called overfitting and clearly requires the availability
of a large quantity of data. The second approach is to start with an
appropriate simple model. The forecast errors, or residuals, are then
carefully studied to see whether they show any remaining structure, e.g.
autocorrelation or periodic behaviour. If such a structure is found and a
suitable model fitted to it, then a combined model can be created.
Methods for the analysis of errors and, equivalently, residuals are
183
discussed in Chapter 15. If no structure is found in the errors, then our

chosen model can be taken as satisfactory. Thus we have two
approaches, one based on breaking down a complex model, the other
on building up a simple one.
12.3 Model testing*

Forecasting is essentially a dynamic activity and one is usually
concerned with questions as to whether the method currently used is
still successful. Problems of this nature are about the quality of one’s
forecasting and are termed quality control problems. These are
discussed at some length in Chapter 16. Often forecasting methods are
applied almost as soon as any data are available and the problem of
identifying and testing some suitable model is tackled on the basis of
making, one hopes, inspired guesses and then applying quality control
methods to test whether one is right. A consequence of this is that
methods for formally testing models on an adequate set of data are not
used in statistical forecasting as frequently as they are in other branches
of statistics. None the less, if one has a set of data whose structure is
not obvious, then there is a wide range of tests that one can use to test
the fit of the more common models. These tests are also appropriate to
a study of forecast errors. If forecasting is successful, the forecast errors
should be a purely random sequence. If there are any correlation or
seasonal features left in the error sequence, then it should be possible to
improve on the forecasts. The aim of this section is to indicate briefly
the nature of some useful tests. It will be assumed that the reader is
familiar with the main features and forms of significance tests. Almost
all introductory texts in statistical theory cover significance tests.
There are three main areas in which such tests might be useful to the
forecaster.
(a) Testing for randomness
Here the question is whether there is any structure in the data at all.
Kendall (1973) gives a number of different tests of randomness. Table
12.3 describes and illustrates one such test.
(b) Testing for trend
Table 12.4 illustrates two approaches to testing for a trend in the

data. In part (a) a trend model is fitted and the test is whether the slope
is significantly different from zero. In part (b) the data is used directly
to derive a suitable test statistic. Notice that here the test statistic is
designed to be particularly sensitive to trends. Thus in the example a
trend is detected and the hypothesis of randomness rejected, whereas
\
184
Table 12.3 A turning point test of randomness
1. Definition
A turning point is an observation that is greater than both its neighbours or
less than both its neighbours.
2. Examples
+ is a turning point.
+
+
• • •
+ + + •• • *
• • +*® + +•
#
e + .
• •
• •
+ * + + + *
(a) Random sequence ( b ) Oscillating sequence
•
•
•
• ••
( c ) Trending sequence
3. Formulae
n = number of observations
p = number of turning points
pp = expected number of turning points = \(n — 2) for random sequence
op = standard deviation ofp = y{ (16n — 29)/90}
u = (p — Pp)lop = standardized value of p
4. Test
If data are from a random sequence u has, approximately, a standard
normal distribution. Thus reject hypothesis of randomness if | u | > 1.96
for 5% significance test.
5. Example
Data 9.1 8.3 7.2 10.3 9.5 10.4 10.5 10.1 9.7 10.6
Turning + + + + +
points
n = 10, p = 5, pp = 5.3, op = 1.2, u = 0.28
The observed u is near the centre of the standard normal distribution so
the data is consistent with the hypothesis of randomness.
185
Table 12.4 Two tests of trend
(a)
1. Formulae (see Chapter 5 for detailed notation)
B = least squares estimator of slope (/3)
o = estimated variance = 2e2/(n — 2)
6j = estimated variance of /5 = d2/2C2
£=(/?— 13)1 dp = standardized value of (3
2. Test
If there is no underlying trend in the situation, j3 = 0. The statistic t = fildp
can then be tested using a f-test.
3. Example (from Table 5.2)
n = 11, p = 1.882, d2 = 2.546, St'2 = 110, = 0.023, t = 81.8, which is
clearly highly significant, i.e. there is a clear trend in the data.
(b)
1. Formulae
P = number of pairs of observations, not necessarily adjacent, in which
the later one is greater than the earlier one
lip = expected value of P = \ n(n — 1) for random sequence
4P f +1, steady upwards trend
r= “1; r =
n(n — 1) T, steady downwards trend
or = standard deviation of r = ^J{2(2n + 5)/9n(n — 1)}
2. Test
If data are from a non-trending random sequence, r/or has, approximately,
a standard normal distribution. Thus reject hypothesis of non-trending
randomness if | r/or | > 1.96 for 5% significance test.
3. Example
Data 9.1 8.3 7.2 10.3 9.5 10.4 10.5 10.1 9.7 10.6
Number 0003 3 56449
of points
to left
with lower
value
P= 34, r=0.51, or = 0.25, r/or = 2.05
The observed value r/or is just significant at the 5% significance level; so
there is some evidence of a trend, but more data would be useful before
reaching a firm conclusion.
186
Table 12.5 An analysis of variance test of seasonality
(a) Form of data and calculations
Season j
1 2 r
1 Xi x x x
12 lr
2 x x x
21 22 2r
Period
t X
tl x
t2 x
tr Totals
Totals Si s 2 Sr XSj
Sum of squares SSi ss 2 ssr ZSSj
2 2
(e.g. x\ L +X 21 + ...+ x n)
Number of periods data HI n2 nr N

(usually = t) S]/n, S2/n2 S2/nr ZSj/nj
Sum of squared deviations SSDX SSD2 SSDr XSSDj
(i.e. SSD = SS-S2 In)
Degrees of freedom nx — 1 n2 — 1 nr — 1 n—r
• M
&
.*1
Notation = Sj/nj, SSj =

%xij.
ii
•s*
i i
X' = Xxj/r
i
Variance analysis
Total sum of squares

9 (
K2S,)
2
22(*,y - x.f = ESS,- - —y- (1)

i j A
, , (ES,)2
Between season sum of Xrij(xj — x) = XSj/rij — (2)
squares
Within season sum of 2E(Xy - Xj)2 = ESSDy = (1) - (2) = (3)

squares i i i
(b) Analysis of variance table and test
Source of Sum of Degrees of Test

variation squares freedom (S of S)/(d of f) statistic
Between (2) r—1 Sl-(2)l(r-l)

seasons
F = S2b/S2v
Within (3) N~r =
(3)/(N — r)
seasons
Total (1) N- 1
'j o
If there is no difference between the seasons, then S[j/Sw has an

F-distribution with (r — 1, N — r) degrees of freedom. Thus if F
is significantly larger than one, we reject the hypothesis that
there is no seasonal effect.
187
(c) Example: forecast errors (rounded) in quarterly forecasts
Quarter
1 2 3 4
Year 1 -10 2 9 -6
2 -3 -1 4 0
3 -4 5 1 -3
4 2 2 -6 -8
5 -2 4 7 4 Totals
Totals -17 12 15 -13 -3

Sums of squares 133 50 183 125 491
Number of years 5 5 5 5 20
S2/n 57.8 28.8 45.0 33.8 165.4
SSD 75.2 21.2 138 91.2 325.6
Degrees of freedom 4 4 4 4 16
(-3)2
Total sum of squares 491 - 490.55
20
(-3)2
Between season sum 165.4 164.95
of squares 20
Within season sum Difference 325.6 checks with

of squares SSD above
Source of Sum of Degrees of

variation squares freedom (S of S)l(d of f) F
Between 164.95 3 54.98

seasons
2.70
Within 325.60 16 20.35
seasons
Total 490.55 19
From the tables the value of F needs to be 3.24 to be

significant at the 5% significance level. Thus there is no
clear evidence that there is any seasonal variation left
in the forecast errors.
for the same data the test of Table 12.3 was unable to detect
non-randomness.
(c) Testing for seasonality
Table 12.5 illustrates how the method of analysis of variance, which

is dealt with in many statistics texts, can be applied to testing for
seasonality in data.
12.4 The use of judgement
There is a tendency for those, such as the writer, who approach

forecasting from the statisticians’ viewpoint to disapprove somewhat of
‘judgement’ as a means of obtaining a forecast. However, as we
admitted in Chapter 1, even the most statistical of statistical forecasts
involves the assumption of an underlying stable structure, at least
locally. Whether this is so is a matter of judgement. It is a judgement
based on a study of the factors in the situation. One might, for
example, list the factors in the situation, government policy or
consumer attitudes say, and judge whether they are likely to change
sufficiently in the lead time to upset one’s model. If the answer is yes,
then one may have to produce a forecast by applying one’s feel for the
situation to adjust the mathematically derived forecast. The point,
however, is that even if one leaves the statistical forecast alone that
decision is based on judgement.
Forecasts are sometimes produced by judgement alone. Often this is
done because there is insufficient time allowed for obtaining and using
data in a statistical forecast. Sometimes the data may not exist at all or
exist only in inadequate amounts. It may be that a situation is changing
so rapidly that a statistically based forecast would be no use even as a
guide. This does not mean, however, that studying the situation from a
model-building point of view would not help in the forming of
judgement. The very essence of producing good forecasts by the use of
judgement is the careful study of the situation from as many viewpoints
as possible.
On the basis of a feeling that more heads are better than one, many
judgemental approaches are based on getting a consensus forecast from
a number of individuals. These vary in name and type from the
committee of area sales managers and the ‘Jury of Executive Opinion’
(Bassie, 1958) to the surveys of forecasts by relevant experts carried
out by learned societies, banks, newspapers and magazines. There is a
temptation here to argue from the fact that an average, x, is a better
estimate of the underlying mean than a single observation. In fact one
can argue that as some forecasters will be better than others, one should
use a weighted mean of the individuals forecasts, with weights
increasing with the individuals past successes at forecasting. If all the
experts produce unbiased forecasts, this procedure would certainly be
reasonable. However, it is likely that an unexpected downswing in, say,
sales would be unexpected to all the group and the consequent bias
would appear in the consensus forecast as in the individual. However,
this does not rule out the method but rather points out that it has no
magic ingredient. The problem of how to combine the individual
forecasts has been met with a variety of solutions. The simplest is the
type of averaging operation suggested above. This could be modified to
189
allow for the bias that will almost certainly occur in the forecasts of
some members of the group. This occurs because of the way in which
the forecast will influence their part of the organization; a high forecast
will make life harder for production men but possibly more lucrative to
salesmen. Over a period of time the magnitudes of these biases could be
assessed and these forecasts adjusted accordingly.
Other methods seek to get the group to work as a team and arrive at
a genuine consensus. This can be done by circulating initial forecasts,
each with some indication of the reasoning behind its choice, on the
basis of which each member revises his forecast. This is usually done
anonymously. The feedback produced by the study and revision of
forecasts reduces the spread of the forecasts. This is particularly so if
those well above or below the average have to produce detailed
arguments for keeping to their forecasts, that is if they wish to. Part of
the value of this type of exercise is the investigation of relevant factors
as seen from a number of different viewpoints. It is possible in this
technique to introduce something corresponding to the weights in the
previous method. In the initial circular, previous forecasts and actuals
could be given for each of the forecasters. Thus people would begin to
form a clearer assessment of their own and other people’s biases and
tendencies.
A major problem occurs with judgement forecasts that is not present
to the same extent with forecasts based on models. With a statistical
forecast, assuming stability, one can evaluate standard errors for the
future observations, given the forecasts. If the same forecaster makes
repeated forecasts on the same basis of information, then there will
exist a set of errors with probably a clear statistical pattern. Where we
are dealing with consensus forecasts using possibly different people and
different lines of argument each time, it is likely that the errors over a
sequence of forecasts will not show so much stability. If this is the case,
then it is difficult to assess limits for the forecast. One could try and
form a consensus opinion of the range in which the future observation
will lie with given probabilities. Alternatively, the revised, or original,
sets of individual forecasts could be treated as a set of random
observations on a distribution with the consensus forecast as mean. This
distribution could then be used to provide an error distribution.
Experience would have to be built up to see how close the distribution
of errors in forecasts corresponded to the distribution of forecasts made
by individuals.
References
Bassie, V. L. (1958). Economic Forecasting. McGraw-Hill, New York.
Kendall, M. G. (1973). Time Series. Griffin Co. Ltd. London.
.
Part III
The forecasting process
The following chapters explore a range of problems and methods

associated with the development and implementation of statistical
forecasting methods.
•S
Chapter 13
Data
13.1 Introduction
The quality of a forecaster’s results cannot be better than the quality of
his data. It is therefore worthwhile to devote a short chapter to looking
more carefully at the forecaster’s data. In section 13.2 we will look at
the sources of his data and in section 13.3 we will examine the quality
of his data. The ‘quality’ of data is almost impossible to measure or
define, so we will avoid the issue by simply looking at situations where
the ‘quality’ of the data is clearly low and seeing what might be done
about it.
We will also look at some of the many ways in which data are
adjusted for use in forecasting.
13.2 Sources of data
The sources of data depend on the type of variable being forecast. The
sources for forecasting sales of furniture will be very different from
those required for forecasting the number of school places needed for
ten-year-olds in a town. There is no hope, nor need, to give a study
covering all applications. Instead, we will concentrate on forecasting in
the business area of application, as this is the most common and as it
also illustrates a number of more general aspects.
In Chapter 2 we made a distinction between internal and external
data, i.e. data obtained within the forecaster’s own firm and data
obtained from outside. Let us look at internal data first. It is essential
to the forecaster’s work that, as part of the firm’s procedures,
appropriate data are collected and recorded. He should look very
carefully at how data are initially obtained, combined and circulated so
that he can obtain data of the highest quality possible as early as
possible. Where the forecaster is starting from scratch and trying to get
together past data, a lot depends on knowing the firm and its staff very
193
\
194
well. In one firm the sales department discovered that, unfortunately,

they could not trace any records of an important piece of information
about orders for any time before the previous year. This was casually
mentioned to the production manager who thereupon produced a sheet
of paper on which he had noted this information every month for the
last ten years Tor general interest’.
When handling the firm’s own data it tends to be assumed that the
firm knows all about it. Where the data originate from the ‘customer’,
there are often many aspects that are not apparent from the numbers
themselves. What is the customer buying for? What is his buying policy?
These are very relevant questions. A customer who buys material may
change his use of the material and thus possibly his future buying
pattern. A customer who buys both for direct use and for stock may
well produce misleading orders in times of change. The more the firm
knows about its customers the better. Where there are a few major
customers, relevant information may be provided willingly by the
customer if he can see that it could benefit him in terms of, say, more
reliable deliveries. Alternatively, he could be persuaded to provide
information by way of appropriate offers. Where there are a lot of
customers the problem is more complex. The natural procedure is to
use samples. This, however, is very expensive, but some consulting firms
use such samples to study large numbers of products and can thus
spread the costs. The information on buying patterns, etc., provided by
such firms can often be very valuable, particularly in relation to new
products and to assessing who buys the products.
Forecasters in the past have often been defeated by the lack of data.
The current developments towards the creation of effective data bases
for firms should greatly ease the forecaster’s problems in this context.
As well as internal information, an effective information system for the
forecaster should contain information about the external environment.
Pearce (1971) gives a study of such systems within the marketing
context.
As well as the problems of data shortage, the forecaster is sometimes
faced with problems of data abundance. In a stock control situation of
a manufacturing firm there may well be many thousands of items
requiring a weekly forecast. In such a situation a detailed study of
models and methods for forecasting each item is out of the question.
Probably the best procedure is to seek to classify the items into groups
on the basis of criteria such as type of use, value per unit and
magnitude of use. A random sample of items from each group can then
be used as the basis for examining forecasting methods. It would be
sensible in this situation to use some items in each group as the basis for
choosing methods and parameters and other items for checking the
effectiveness of these methods. Clearly, the robustness of the methods
195
is of great importance when they are used to forecast series that have
not themselves been investigated.
External data for business forecasting come under roughly two
headings:
(a) data for similar products, or generally for the same industry;
(b) data concerning the general business environment and economy of
the country.
Considering industry data sources first, these are mainly the

institutions or societies representative of the industry. Some industrial
magazines publish a wide set of statistical data on sales, etc., for the
industry as a whole. These data frequently come from government
sources and are based on returns made by individual firms to central
government. How much detailed statistics has to be supplied to central
government by firms varies from country to country. Some industries
have schemes by which firms can join an organization to which they
supply confidential information on many things from sales to wages.
This information is then compiled in such a way that the data from
individual firms cannot be identified and the statistical report is sent on
a monthly basis to all firms in the organization. This approach enables
firms to have useful information very quickly. Mention should also be
made of commercial consulting agencies and research organizations who
are either associated with particular industries or make occasional
studies of these industries. These organizations often produce statistics
on an industry of relevance to forecasting.
National data provide information to assist in the general judge-
mental background of forecasting as well as in the statistical techniques.
These data include statistics on the economy as a whole, on particular
parts of the economy, including data on geographical regions of the
country. As well as this general information, some agencies select and
publish series that are particularly useful as lead series, or as indicators
of turning points and cycles. As all these publications vary from
country to country, there is little point here in listing large numbers of
publications. In the books by Silk and Curley (1970) and Woy (1965)
the reader should find information to indicate how to look further.
13.3 The quality of data
Let us turn now to the quality of the data used in forecasting and list
some of the things that can be wrong with such data.
(a) The data may be unavailable or late. For example, in regression
methods substitute variables have to be used where an intuitively
important regressor variable is either not known or is not published
until after the deadline for producing the forecast. The substitute
\
196
variable may be an estimate of the missing one based on sample

information or it may be some other variable that has a high
correlation with it. There are occasions when information can be
speeded up. Certainly government statistical agencies have been
working hard in this direction in recent years. In a number of cases
firms have greatly improved their very short-term forecasting for
stock control use by replacing mailed information by telex or by
remote computer terminals.
(b) The data may be inaccurate. Before arriving on the forecaster’s
desk, the data will have been collected and processed in various
ways. What can be done about inaccuracy will depend on the
situation. A number of examples will suffice to point to the value
of looking at this aspect of the data. A firm based its forecasts and
its long-term production schedules on annual estimates by its
customers, who were large and few in number. Originally the data
produced bore only a remote relation to the final orders. At a later
stage these were significantly improved by offering the forecast
amount at a discount. The firms therefore found it worthwhile to
think more carefully about their annual estimate. A similar
procedure has been adopted with sales forces where financial
incentives have been introduced to encourage accurate estimation
by the salesmen of their future sales. Where data are copied in
transmission, errors are almost certain to creep in. Where forecast-
ing is done by computer on a large scale, it is worthwhile including
in the programmes simple tests to check that numbers have the
right order of magnitude and the right form.
13.4 The adjustment of data
When one talks to practitioners of forecasting it is often apparent that,

though they will say that they use some standard method, they have
their own particular versions appropriate to their own particular
situations. The most common variations are linked with the idea of
making adjustments of various types to the data or the forecasting
formulae. In the first of these types the raw data are altered in some
fashion before being put in the forecasting formula. In the second, the
raw data are used but adjustments are made to the formula to allow for
some features of these data. The aim of this section is to briefly discuss
and reference some of these adjustment procedures.
13.4.1 Adjusting for time period
Most forecasting methods are based on assuming that data arrive at

fixed regular intervals. This assumption may not be valid and the
adjustment is made to produce a new set of data for which the
197
assumptions are more nearly valid. For example, in a regression

problem, the regressor variables may be yearly figures. If one is
published quarterly, then clearly the quarterly figures can be added
together to obtain annual figures. The problem, of course, comes when
the situation is the other way round. Quarterly figures may be
obtainable for all the variables save one, which has only annual data. In
this case it may be necessary to forge quarterly figures for the variable
in question. The only necessary condition on these figures is that they
add up to the correct yearly totals, xt. One approach is to make each
quarterly figure a weighted average of three yearly totals. Thus the
quarterly figures for year t, y\, y]1, y]u, y]v, are given by
y\ = axt_ x + ext + dxt+ x
y\l = bxt_ j + fxt + cxt+!

y111 = CXt_ ! + fxt + bxt+ !
y|v = dxt_ j + ext + axt+ ,
It will be seen that the weights a, b, c, d, e, f are chosen to give a

reasonable symmetry to the weighted averages. To choose actual values
for a, b, c, d, e, f, we add to the above condition some further
reasonable conditions. For example, if the quantities xt_x, xt, xt+1
follow a trend, y\, y\l, y\u and yjv ought to follow the same trend.
Boot, Feibes and Lisman (1967), Ginsburgh (1973) and Lisman and
Sandee (1964) give a fuller discussion with some appropriate formulae.
Another area of common occurrency is that produced by the use of
unequal time periods. For example, a firm’s monthly figures may be
based on four five-week months and eight four-week months. To
produce a set of comparable figures, we must adjust each monthly figure
by multiplying by a weighting factor, say w1,if there are four weeks in
the month and w2 if there are five weeks. To obtain comparable
figures, if weekly sales are constant, say one unit per week, we must
have
w x • 4 = w2 * 5
We must also have a set of figures giving the correct annual total, so
that
8 • IUJ • 4 + 4 * w2 * 5 = 52
Solving these two simultaneous equations gives
wx = 13/12 and w2 = 13/15
Having multiplied each actual monthly figure by wx or w2 and
substituted the adjusted figure in the forecasting formula, we must
\
198
remember to divide the forecast by wx or w2, as appropriate, to get a

proper forecast of the future value.
In the above situations we have had some basic time interval from
which to make adjustments. It can occur that the data arise at
completely irregular times tx, t2, . . . tk. The observations xx, . . . , xk
would usually represent the number of ‘items’ arising in the intervals
(0,tx), (tx ,i2 (tk_ i ,tk), respectively. In this situation the under-
lying variable of interest is the rate variable, the number of items arising
per unit time. We would thus adjust each observation by dividing it by
the length of its time intervals. The adjusted data, the rates, would thus
be rx, . . . , rk, where rk = xk /(tk —tk_x).
As forecasting methods are often based on discounted methods
(undiscounted methods being included here as a special case with
weights w = 1), we can use this type of data in the standard methods
provided we use appropriate weights for the discounting, for example,
if tk — hj is a time in the middle of the interval ( tk_}_ x, tk_j), then
the weight associated with rk _y- could be ah k This would modify the use
of exponential weights in a suitable fashion. An alternative procedure is
to use standard forecasting formulae modified in an appropriate
manner. By way of an example, consider the modification of
exponential smoothing. The basic formula is
rk = ark_l + (1 — a)rk
We now seek to modify this to allow for the irregular intervals. We

would expect that rk represents a better estimate of the rate for larger
time intervals tk — tk_ x = T, say, so that a choice of a that varied
inversely with T would seem appropriate. One way of picking out such
a relationship would be to vary a in such a way that the r k values have
some convenient property. One suggestion is that if the rate follows a
linear trend the lag in r remains constant in spite of the variation in T.
Suppose that the rate has a slope a and is p at time tk. The lag of ?
behind r will be proportional to a and denoted by ca. From the
diagram of Figure 13.1 it is seen that the expected values of rk and
rk _ L are given by
E(rk) = n~ca
E(rk_ x) = iu—aT—ca
The expected number of items arising in the time T will be the average
rate times T, i.e. (p — aT/2)T. Thus the expected value of rk is
E(rk) = II-OLT/2
Thus if we substitute these expected values in the exponential smoothing

equation, we obtain
li — ca = a(ii — aT — ca) + (1 — a)(p — aT/2)

199
Figure 13.1 Trending rates
Solving gives a = (2c — T)/(2 c + T). Thus at each use of the recurrence
relation a is recalculated from T. The choice of the best a is replaced by
the choice of the best c.
A final comment on the consideration of time scales is that even if
one can obtain data at regular intervals it is not necessarily the best
thing to do. As an example of this point, consider some attempts that
have been made to model share prices (Mandelbrol and Taylor, 1967).
The obvious time scaling is to take one unit as one day’s trading, so that
the xt are the closing prices on the Stock Exchange. However, some
days are slow days for a given share, with very little buying and selling.
On other days as much stock is bought or sold as in the average week.
On these latter days it is as though time speeded up for that particular
stock. It is reasonable to try to scale t so that a uniform rate of buying
and selling occurs. This is done by observing xt at times called t = 0, 1,
2, . . . , which correspond to the moments at which, say, 0, £10,000,
£20,000, £30,000 pounds worth of stock has been traded, provided this
information is available. This keeps a constant amount of trading,
rather than a constant amount of physical time, between successive
observations xt.
13.4.2 Adjusting for known causes
In many sets of data there are certain effects that are known to
influence the raw data. Examples are
(a) holidays, as in sales from retail firms;

(b) the weather, as in sales of ice-cream;
(c) seasonal factors generally;
(d) strikes and other known hold-ups;
(e) stockouts.
Ways of dealing with these vary greatly. Some are pure intuition, where,
\
200
for example, a guessed amount is added to allow for holiday lost sales.
Others are fairly sophisticated methods, where, for example, parameters
are introduced to correspond to the known factors, which are
estimated from the data and then used to adjust the data. The latter
methods are clearly preferable if one is dealing with repeated
occurrences. In unique situations caused by, say, a strike, one is forced
to make a guestimate on whatever grounds seem reasonable. As an
example of the more mathematical approach, a model for ice-cream
sales was expressed as:
=
Xj,k + Ptj + 7r/ +
5Tj,k + ejk
where Xjk is the sales in week j of year k, tj is the temperature norm
for that week, r;- is a seasonal term and Tj k is the average deviation of
that week’s temperature from the norm. A set of past data was used to
estimate the parameters. A figure for adjusted sales for normal
temperatures was then found by subtracting the estimated effects of
Tj k, namely $Tjfki from Xjk. Another common example is the use of
seasonally adjusted data for regressor variables in regression forecasting.
One general comment here is that though such adjusted data are often
useful at the data analysis stage, they may not be so useful for
forecasting. To see why this is so, note that the process works in three
stages:
(a) A model for the data, involving the effects of interest, is fitted.
(b) The data is adjusted by eliminating these effects.
(c) The adjusted data is used with an appropriate forecasting model.
These three stages can usually be combined by incorporating the model
for the data, with the effects, in the general forecasting model.
Experience indicates that this more general model will usually provide
better forecasts than the three-stage approach.
13.4.3 Transforming the data

A very common practice in forecasting is the application of simple
transformations to the data before carrying out the forecasting. The
two most common transformations are
yt = loge Xf
yt = xt —xt_ !
The first of these transformations is commonly applied to reduce
percentage changes to absolute changes, to reduce a growth model to a
linear model or to change from a multiplicative to an additive seasonal
model. The forecast of xt+ t is obtained as ey’t+ 1. As mentioned in the
discussion of growth curves, an unbiased forecast of yt+ j will lead to a
201
biased forecast of xt+1 when this transformation is used. It is possible

to make allowances for such biases and this is discussed in section 16.2.
The second transformation is frequently used as a device for
removing the effects of a wandering mean. If the mean pt is locally
constant, then the differenced series yt will not depend on jiq, since it
cancels between xt and xt_ x. The forecast xt+ 1 will be xt + yt+ x.
There are obviously many transformations that might be useful in
particular circumstances. Two questions need to be considered in
deciding whether to use a transformation. The first is the obvious one
as to whether better forecasts of x are obtained by the use of a
transformation. The second question is whether there is a reasonable
intuitive or theoretical justification for the use of the particular
transformation. If we are to expect the transformation to lead to good
forecasts in the future, we must have some reason to believe that it
related to the underlying structure of the situation.
13.4.4 Adjusting for outliers

Most practical forecasting systems contain quality control procedures
that will pick out values that are in some sense extreme. For example,
the forecast error may be more than say 2lA standard deviations from
the mean error. Such values are often referred to as outliers. The
quality control procedure will examine such items to see if they are
indicative of a real change in the underlying situation. Often, however,
they are just isolated occurrences. In initial investigations of forecasting
it is advisable to put some effort into seeking explanations for outliers.
For example, in forecasting for production planning purposes a number
of outliers were found to correspond to single exceptional orders which
needed special production arrangements. The procedure was not
therefore to try and forecast the outliers along with the rest, but to
separate them as a special category. The result of this was to improve
the quality of the forecasts made for the normal run of items.
Where outliers occur in an operational forecasting system, it is
advisable to replace the outlier by its forecast or some other suitable
less extreme value, otherwise future forecasts based on the outlier will
be badly out. An indication of less subjective approaches to outliers,
transients, was given in section 10.3.
References
Boot, J. C. G., Feibes, W., and Lisman, J. H. C. (1967). Further methods for the
derivation of quarterly figures from annual data. Applied Statistics, 16, 65—75.
Ginsburgh, V. (1973). A further note on the derivation of quarterly data consistent
with annual data. Applied Statistics, 22,368—374.
Lisman, J. H. C., and Sandee, J. (1964). Derivation of quarterly figures from annual
data. Applied Statistics, 13, 87—90.
202
Mandelbrol, B., and Taylor, H. M. (1967). On the distribution of stock price

differences. Opns. Res., 15, 1057—1062.
Pearce, C. (1971). Prediction Techniques for Marketing Planners. Assoc. Bus.
Programmers, London.
Silk, L. S., and Curley, M. L. (1970). A Primer on Business Forecasting. Random
House, New York.
Woy, J. B. (1965). Business Trends and Forecasting—Information Sources. Gale
Research Co., Detroit.
Chapter 14
Adaptive methods and other extensions
14.1 Introduction
We have discussed thus far a fairly limited number of forecasting
methods. The practicalities of forecasting often pose problems that are
not adequately dealt with by these methods. The situation may change
so rapidly that even our local model-building approach will not work. A
‘fracture’ may occur in which a model parameter suddenly changes its
value. The situation may be highly complex and involve variables other
than time. It may require the use of different approaches at different
stages in the problem . It may be that such new problems will require
new models and new methods. However, experience suggests that it is
worthwhile to look at the methods that we already know and seek to
modify and to extend them to meet the new situations. The aim of this
chapter is to introduce some of the ways in which this is possible.
14.2 Adaptive methods

We consider in this section ways of getting our standard forecasting
methods to adjust to rapid changes in parameter values. Clearly, if
global methods are being used then, when a change has been found to
occur, we must wait until enough data have been accumulated to enable
a model to be refitted. Thus global methods are not very suited to this
type of situation. Local methods are more suitable but they need to be
adjusted to deal with rapid changes.
To see how this can be done we take, as our illustration, forecasts
based on simple exponential smoothing as discussed in Chapter 4. On
several occasions in past chapters we have observed that decreasing the
smoothing constant, a, makes a forecast more sensitive to variation in
the data, both random variation and that due to real changes in
structure. Conversely, increasing the smoothing constant reduces the
error variance, but also makes the forecast slower at following changes
203
\
204
in structure. It is therefore natural to look into ways of changing the

smoothing constant while forecasting, the general aim being to reduce it
on occasions where there are changes in mean and to increase it when
the situation is stable. There are several ways of doing this that have
proved effective in practice. One possibility is that of using a subjective
assessment of the stability of the situation as the basis for adjusting the
smoothing constant. In many situations this is obviously a sensible
approach. For example, if a large advertising campaign is to be
undertaken, one would hopefully reduce the smoothing constant so
that the forecasts would rapidly pick up any increased level of sales.
To develop some more objective ways of varying the smoothing
constant, consider first how a single fixed value might be chosen. First
we choose a set of alternative values for the smoothing constant a, e.g.
0.2, 0.4, 0.6, 0.8, 0.9, 1.0. Applying each of these in turn, we can
calculate forecasts for our data and for each set of forecasts evaluate
the appropriate forecasting criterion C, e.g. C = MSE. We finally choose
the value of a that gives the smallest value of C. This procedure can be
modified to enable the value of a to be adjusted as forecasting proceeds.
We may adjust a at regular intervals by going through the above
procedure each time, using only the data in the last interval (e.g. see
Eilon and Elmaleh, 1970). Alternatively, we may re-evaluate a, using
only recent data, whenever some quality control procedure indicates
that a change may have taken place. Ideally we would like to be able to
adjust a, if necessary, every time an observation is taken. One way of
doing this is to evaluate some local forecasting criterion at only a
limited number of values of a. We might, for example, fix on two values
of a: a high value for stable situations and a low value for periods of
change. The high value would normally be used, but the low value
would be dropped to whenever the value of C indicated that it was
necessary. Another approach which continually searches for the best a,
but only requires the evaluation of C at three values of a, is illustrated
in Table 14.1. Details of these and related methods are given by Chow
Table 14.1 A search for the best a
Values of C Values of a used Best value

for given a 0.70 0.75 0.80 0.85 0.90 of a
75 63 79 0.80
73 68 81 0.80
71 70 80 0.80
70 72 81 0.75
75 69 73 0.75
Time 77 69 73 0.75
205
Figure 14.1 Forms for 7(e)
(1965), Roberts and Reed (1969), Roberts and Whybark (1973),

Montgomery (1970) and Rao and Shapiro (1970).
Yet another approach is to make the smoothing constant a function
of forecast errors. The method of Van Dobben de Bruyn (1964) does
this by using simple exponential smoothing in the error correction form
xt = xt_1 +y(et)et
The smoothing constant is thus made a function of the latest error et.
By choosing the form of the function y(et) to be like that in Figure
14.1, it is possible for the formula to behave like ordinary exponential
smoothing with a high smoothing constant when et is small. When,
however, large errors occur, the value of y(et) increases and these errors
have a consequent large effect of the estimated mean. A danger here, of
course, is overreaction and a consequent instability. The original work
of de Bruyn allows for this by using a more elaborate form of the error
correction approach. An alternative is to make the smoothing constant
depend on the whole sequence of past errors, though obviously with
emphasis on the latest values. We thus need to devise a function that
either increases or decreases when the sizes of errors tend to get larger.
One such function, that will be discussed further in Chapter 16, is
defined by
| St | = | smoothed error/smoothed absolute error |
etlI et ||
The smoothed error gives a measure of bias in the errors. The smoothed
absolute error gives a measure of variability with which to compare the
bias. It can easily be shown that 0 < | S |< 1 and it is clear the | S | will
equal one when all the errors have the same sign. The smoothing
constant can thus be set equal to 1 — | S | or to some multiple of this,
possibly with a restriction to limit | S I below some maximum value.
The restriction is sometimes introduced to stop the smoothing constant
getting unrealistically small. Thus one might have
1 — c | St I, | St I < S max
1 — c S max, | St I > S max
\
206
The use of such an ‘adaptive’ parameter was introduced in Trigg (1964)

and Trigg and Leach (1967) and the idea has been widely used e.g.
Bamber (1969). One word of caution is necessary for the case where
there are several parameters in the model. If the estimators of
parameters such as slope are allowed to change rapidly owing to the use
of an ‘adaptive’ smoothing constant, the forecasting method can
become unstable and give wild forecasts. In these cases it is best to use a
form of forecasting in which several smoothing constants can be used.
That associated with estimating the underlying mean of the process can
be made adaptive, but the other smoothing constants should be kept
constant. By way of example, for forecasting a linear trend, Holt’s
method, as discussed in Chapter 5, would become
fit = fit_ j + ftt- i + (1 ~ at)et

+ ce
fit ~ fit- I t
where at is obtained adaptively and c is a fixed smoothing constant.

The approaches above all adapt to changes in the model parameters.
A more general alternative is to seek to adapt to changes in the model
itself. In this approach the data obtained to date are used to select the
most suitable model from a particular set of possible models. As
another approach it is possible to consider models in classes and to
always use the simplest model in the class that currently fits the
situation. For example, the constant mean model is a special case of the
linear trend for the slope (3 = 0. If in Holt’s method, Chapter 5, we
constantly update ft but use the constant mean forecast unless ft is
significantly different from zero, then we can at least be adaptive
between these two models.
14.3 Extensions to recurrence formulae

When exponential smoothing was first discussed, it was shown that
when ar can be neglected the exact form can be approximated by the
following reccurrence relation:
xt = (1 — a)xt + axt_ !
or, in words,
New forecast = (1 — a) x new observation + a x last forecast
In section 5.2 we introduced an extension of the recurrence form to

allow for a trend. This method was called Holt’s method. It will be
recalled that to use this we denote the current mean by and the slope
by fit. The model was thus
Xf— yi ~ fith + h
207
the recurrence relations for mean and slope were
(it = (l — a)x, + a(A,_ , + 0t_ |)
and
ft = (1 — b)(A, — A,_ i) + bft- i
or, in words,
New estimate of mean = (1 — a) x estimate based on new data

+ a x estimate of current mean based on last
estimates of p and (3
New estimate of slope = (1 — b) x estimate of slope based on change
in estimated means
+ b x last estimate
The use of two recurrence relations introduces the use of two
smoothing constants a and b.
The aim of this section is to illustrate how the use of recurrence
relations can be extended to deal with more complex situations. For an
example we consider forecasting sales in a situation where weather is a
factor of importance. This particular feature occurs for many
commodities, from ice-cream to electricity. The papers by Berrisford
(1965) and Chen and Winters (1966) give specific examples of this type
of analysis. Suppose that sales xt for day t can be modelled by a
constant underlying mean nt to which is added the effect of
(a) the deviation dt of the temperature on that day from the norm for
that day in the year, the norm being given by the average
temperature over a number of years for that particular day in the
year.
(b) the day in the week effect Wt(i = 1, . . . , 7), since, for example,
sales on Saturday, i = 6, may be made higher than those on any
other day.
The final model is thus
xt = fJ-t + Wi + adt + et
where a is a parameter, that we will assume known, measuring the
increase in sales per degree deviation in temperature. In practice a
would probably vary according to the time of year. As usual et
represents the random variation in the situation. Suppose at day t we
have estimates of nt and Wh denoted by jlt and Wt, and also a weather
forecast of dt+l9 denoted by dt+l. The forecast of xt+ i is then
xt+ i = /if + Wi+1 + <xdt+i

\
208
When xt+1 and^dt+1 have been observed, they may then be used to
update jlt and Wt+ i using recurrence relations. Consider juf first. From
the model it is seen that if we have last week’s estimate of Wi+l,
denoted by Wi+ i id > a new estimate of p is given by
%t+ l î+l,old &df+i
Thus this new estimate, based on a single observation, can be put in a

recurrence relation with the old estimate to give
a being the smoothing constant. This formula, as with all our previous
recurrence forms, enables us to re-estimate pf in a very simple fashion,
and also, being of exponential recurrence form, it will tend to follow
local changes in p?. The same approach may now be used again to
update any of the day of the week effects. Thus from the model,
remembering that p?+ x is now known, the quantity
x
t+ I Mr+ I otdf+1
gives an estimate of IT/+1 based on the observed xt+\. Thus again a

recurrence relation for the W values can be constructed to give
^/+l,new ~ (1 )(A? + 1 Mf + 1 &df+ i) + bW i+ i; o Id

where b is a second smoothing constant. If it is thought that o: is a
constant, without any seasonal change but with possible local variation,
we could extend the above to produce a recurrence relation for a,
namely
— +
^t+ 1 (1 1 i Wi+ l ,new )/df+ l C0it
where c is yet a third smoothing constant, which should be very close

to unity.
In our initial introduction of the use of recurrence relations in
forecasting we carried out three steps:
(a) Estimate the model parameters, say by discounted least squares.

(b) Express the estimators obtained in recurrence form.
(c) Use the estimated parameters in the model to obtain the forecast.
An obvious development is to try to reduce the number of steps. In the

example above we have in fact done this, for we wrote down intuitively
recurrence relations for the estimators ju, W and 5. We did not derive
these estimators from any principle of estimation. In practice this is
frequently done. It can sometimes be shown that the estimators
obtained are in fact the same as would have been obtained from steps
(a) and (b) above. In general, however, it is difficult to say much about
the properties of such estimators. A detailed study of a certain type of
recursive estimator is given by Albert and Gardner (1966).
209
Brenner, D’Esopo and Fowler (1968) examine the properties of sets

of recurrence relations that give recursive estimators for parameters in
polynomial prediction. For example, for a quadratic model they
consider the set of relations
+
at = axt (1 ~ a)(S*_ i + fat-i)
ft = b(at — at_ j) + (1 — 6)(ft_ , +7t-i)
—
7t = c(ft ft- I ) + (1 — c)7 f_ j
with the forecasting formula
ft(ft + 1)
- + ft ft + It
1.2
A further step is obviously to try to go directly to step (c), or rather
to go directly to a forecasting formula without reference to model
parameters at all. One approach is to specify the form of forecasting
formula and study its properties. For example, Ward (1963) shows that
the recurrence relation
xt = ooct + fixt_ 1 + (2 — a)xt_ ! — (1 + (3)xt_ 2
includes several standard trend predicting formulae as special cases. He

then examines some of the properties of forecasts obtained by such a
formula. Yet another approach is to start with a very general
forecasting formula and choose its parameters to get the best forecasts
irrespective of the underlying model. We consider in section 14.5 the
most common way of doing this.
14.4 Extensions to the error correction method — Kalman models*
We showed in Chapter 5 that the linear trend model could be forecast

using recurrence relations for the estimated current mean fjt and slope
ft. We also noted that these relations could be expressed in an error
correction form, i.e.
A
xt+\ =A t +Pt
+ +
jlt =
fat- i fat- i (1 a
)et
=
fat fat- 1 + ce,
where et is the one-step-ahead forecast error and a and c are forecasting
parameters. In section 7.6(b) (iii) we noted that these results could be
obtained by using a recurrence formulation for the trend model itself,
namely
xt = M* + et
+ +
Mr ~ Mr- i fit- 1 7r
= +
fit fit- l ^r
the quantities ef, 7f and being independent random variables. It is
found that many of the structures discussed in previous chapters can be
described by models having such a recurrence form. For models of this
form it can be shown that the least squares and discounted least squares
estimators of the model parameters can be expressed in error correction
forms, analagous to the models themselves. Having illustrated this for
the linear trend model, let us give a sketch of this formulation for two
other structures from previous chapters, namely for seasonal and
growth curve structures.
(a) For the seasonal index models of section 9.2.2 a recursive model of
the additive form would be
X, = fit + 4>t,i + et
fit = fit-i + It
0t,j ~ 4*t— r ,j ^ tj (j ~ • • • > **)
Thus the seasonal index <j>tj for each season, j\ is changing by a

random amount at each season. The recursive formulae for
estimating the mean and the seasonal indices are
=
& M?- i +aet
A A
<Ptj = 4>t- r,j + ciet (j = l, ... ,r)

where a and cy are appropriate constants.
(b) As a further useful example, consider a model analogous to the
simple modified exponential growth curve. In the model for this,
given in Chapter 9, we have a linear model if we can assume the
growth rate to be known. An analogous model here is to introduce
an assumed known growth term into the linear trend model. Thus
we rewrite the linear trend model as
+
xt = Mr et
+
Ab = liT-I kfit- ! + yt
= +
fit kfit- I &t
where k (0 < k < 1) is the parameter that controls the growth.
Here the expected value of fit tends to zero and the curve tends to
an asymptotic value. The corresponding forecast and estimators of
211
error correction form are thus

=
&t+1 At +
fit
fit = Af -1 +kj3t_1 +A1et

+
fit =
kfit- 1 -A 2 et
As an aside, it should be noted that in practice k will usually be an

unknown parameter. This may be intuitively estimated from the
data, for example, by exponentially smoothing the quantities
zt = (*f - At-1 )/&-1
We have now seen three examples of linear models based on
recurrence relations with analagous error correction formulae for
updating the estimates of the parameters. The theory and application of
linear models in recurrence form was developed by Kalman and Bucy
(1961) in the early 1960s and further developed for general forecasting
use in the early 1970s by Harrison and Stevens (1971). The basic form
of the Kalman results are:
(a) xt =a linear function of a set of time-dependent parameters 6t + a
random variable et with zero mean and constant variance
e.g.
xt = 1 x ut + 0 x fit + et
In matrix form we write, in general,
xt = ve* + et
V being a known quantity.

(b) A set of equations that relate the current values of each of the
parameters to a linear function of past values of all the parameters,
with added random variation, e.g.
+
M* = lx lit- I 1 x fit-i +yt
+ +
fit = Ox Mr- i 1 x ft- i 5*
The general form is

et = Let_i 4 MY,
L, M being known quantities and Yt representing the random

variables.
(c) For the above models it can be shown that the forecast having the
smallest mean square error is given by
=
xt+1 VL
\
212
where Bt is obtained by an error correction formula of the form:

6t — LBt_ i + K f(?t
The form for xt+l follows naturally from the equations of the
model. In the error correction formula L0f_! gives the natural
updating of the parameters and Ktet provides a correction based on
the latest forecast error et. The correction multiplier Kt is best
discussed in relation to three possible situations that can arise:
(i) If we have a deterministic model or a stochastic model with
fixed parameters, treated in a global fashion, the unknown
parameters do not change in a random fashion. Thus Y = 0. In
this situation we would expect to get better and better
estimators of the parameters 0 as time goes by. Thus we will
take less and less notice of the errors et. For this to happen Kt
will approach zero as t increases. We have already seen an
example of this in Chapter 4 where, for our global constant
mean model, we can write
~ Ah
Ah =Ah-i
1
fit = fit-i +7 e,
L
Thus the correction multiplier is Kt = 1/t which decreases as t

increases.
(ii) If the values of the parameters 0 are influenced by the
occurrence of the random variables Y, then there is no fixed
sequence of values to which 0f can move. Thus Kt will not
necessarily go to zero. In this situation Kt will depend on the
variance properties of all the random components. It is found
that the values of Kt, in both this case and in (i) above, can
themselves be found using recurrence relations. The detailed
derivation and use of these is beyond the scope of this book
and the reader is referred to Harrison and Stevens (1971) and
Kalman and Bucy (1961) for such detail.
(iii) In (ii) the assumption of constant 0, used in (i), was avoided by
defining a precise model for 0f. The multiplier Kf requires
detailed knowledge of the variance properties of the process
for its evaluation. As an alternative approach we may regard
the situation of (i) as a model to be used in a local fashion. We
could use discounted least squares to derive the recurrence
relations for Kt (see, for example, Morrison, 1969). Alter-
natively, we could simply regard Kt as a forecasting parameter,
as indeed we did in our study of Holt’s method in Chapter 5.
213
In this case can be chosen simply as a fixed, or possibly

slowly changing, quantity whose value is chosen to give the
best forecasts.
(d) The ideas of ‘Bayesian’ forecasting introduced by Harrison and
Stevens (1971) extend the approach indicated in (a), (b) and (c) by
allowing the random variable Y to be in a number of possible states.
The underlying concepts of this extension were illustrated at the
end of section 10.3.
14.5 Linear forecasting formulae

A considerable number of the forecasting formulae that we have
studied in the previous chapters are linear in the past observations. That
is to say, they take the form
+
xt = bx xt + b2xt~ I + . . . brxt_ r+ i
For example:
(a) The forecast based on a simple moving average of r terms is
~ _1 1 1
Xt - ~~ Xt H Xt_ i + ...+“ Xt_ r+ j
r r r
(b) A forecast based on exponential smoothing, using the approximate

form for large r, is
xt = (1 — a)xt + (1 — a)axt_ + . . . + (1 — a)ar~ 1 xt_ r+ j
x
(c) The forecast for the autoregressive model of order r is

xt =
<t>\Xt + 1 +• ••+ <t>rXt-r+ 1
(d) It can be shown that the forecast for any moving average model or
mixed autoregressive—moving average model can be expressed as an
infinite series of past observations. However, in practice it is often
found that the coefficients can be set at zero after some suitably
chosen number of values, with very little loss in terms of
forecasting accuracy. The forecast thus takes the linear form above.
We have always justified the use of such formulae on the basis of
some appropriate model. We may, however, take a different approach
and forget all about model building and simply find the values of the
coefficients, 6j, b2,--',bry for which the formula gives the best
forecasts. Such an approach has a number of weaknesses. It tells us
nothing about the structure of the situation, for it does not require
such knowledge to make it work. It will not necessarily produce the
best forecast possible, for many structures lead to forecasts that are not
simply linear functions in the observations. The best linear forecast,
214
however, may be the overall optimum. Even if it is not, it gives a

forecast which involves minimal knowledge and assumptions about the
underlying situation. Thus for a purely empirical approach to fore-
casting there is something to be said for trying out such linear
forecasting formulae. The aim of this section is to show briefly several
approaches that have been used to find the values of the coefficients,
bx, b2, . . . , bt, which provide the best forecasts.
(a) To illustrate the first method, consider the simplest linear forecast
which uses only the last observation xt to provide a forecast of
^t+ k •
x tik = bx,
As the criteria for the ‘best’ forecast we will use the minimum
mean square error. Thus we wish to minimize
s = E(elk) = E(xt+k — bxt)2

= E(x?+k) — 2 bE(xt+kxt) + b2 E(x2)
If we assume, without loss of generality, that E(x) = 0 and

Var(x) = o2 for all time, and denote the autocovariance of lag ft by
C(ft), i.e. E(xt + kxt) = C(ft), then
S = o2 -2 bC(k) + b2 o2
This expression is minimum when
b = C(k)/o2
Thus the forecasting parameter depends on k, the lead time of the

forecast. To turn this into a practical method we need to estimate
C(k) and o2 from the data. The theoretical minimum value of S
achieved by using this value of b is
Smin =02-C(k)2l02
= O2 (1 — p(k)2 )
where p(k) is the autocorrelation coefficient of lag ft. It will be seen
from this that as the autocovariance increases the mean square
error is reduced.
If we take the general linear forecasting formula and carry out
the same calculations as in the example above, we obtain a set of r
simultaneous equations, the normal equations, for the unknown
parameters b, , br:
brC(0) + br_ ! C( 1) + . . . + &! C(r - 1) = C(r + ft — 1)
&rC(l) + br_1C(0) + . . . + blC(r-2) = C(r + k-2)

• •• ••• ••• •••
brC(r — 1) + ,C(r — 2) + . . . + blC(0) = C(k)

215
This assumes, as in the example, that E(x) = 0 over time, and that
the autocovariances C(i) depend only on the time intervals i
between the x values and not on the actual time. The study of the
above equations for practical use is a major subject, see Wiener
(1949), Whittle (1965) and Yaglom (1962), and we limit ourselves
to making the following points:
(i) Given knowledge of C(0), C(l), etc., there are many computer
programs that can solve such sets of equations. The symmetry
of the equations helps to make this a relatively fast process
unless r is large.
(ii) In practice we must estimate the autocovariances. For large
values of r this cannot be done very precisely. If an
examination of the estimates of C(0), C(l), etc., indicates that
they have some structure, e.g. C(k) = pk, as for one of the
models discussed in Chapter 7, then this can be used and the
problem greatly simplified. It is, however, a move away from a
purely empirical approach.
(iii) Robinson (1967) suggests ways of solving the equations by
using first only one past observation, r = 1, as in our example,
then two, three, etc. These methods are not only efficient but
may indicate when we have sufficient past values to obtain
reasonable forecasts.
(iv) If E(xt) = p, for all time, instead of zero, we simply subtract ju
(or our estimate of it) from every x. The forecast of xt + h will
then be our linear forecast obtained from the adjusted
observations plus p, i.e.
Forecast = n + b1(xl — p) + . . . + bt(xt — p)
= kn + 3ct
where xt is the linear forecast using unadjusted observations
and
k = 1 ~bx — . . . — bt
(v) The above method and normal equations generalize in a natural
fashion to deal with the analogous multivariate forecasting
problem (see Robinson, 1967).
(b) An alternative approach to this problem follows from the simple
observation that the form of the forecast corresponds to a
regression on past observations. The linear forecasting formula gives
the fitted value xt k corresponding to the observed value xt+k. The
regression coefficients bx, b2, • • . , br (plus a constant term b0) can
be chosen by least squares or discounted least squares using the
normal equations of Chapter 6. This enables one to make use of
standard regression programmes that use stepwise regression pro-
cedures to select the significantly non-zero values of b to use. We
\
216
would put the values xt+k as the dependent variable and xti
xt_ j,..., xt_r+i as the regressor variables. This approach is
simpler than that of (a) and has proved to be effective in practice,
(e.g. Newbold and Granger, 1974). A variation on this, discussed by
Wheelwright and Makridakis (1973), uses a recursive method to
find the b values instead of the normal equations.
14.6 Using mixed methods

The aim of this section is to describe briefly a forecasting situation
which has a complicated structure and to show how several different
methods of forecasting can be brought together in such a situation. The
example chosen to illustrate this type of situation is one where it is
required to forecast the number of children in different age groups in
State schools in a given school area. Such forecasts are usually required
for the next five to ten years. The first task in solving any such problem
is the listing of the relevant factors. Clearly in this situation we must at
least take account of the following factors:
(a) the number of children in the various age groups in the area at
present;
(b) the number of births over the past few years, and estimated for
future years, in the area;
(c) the number of new children expected to move into the area from
outside during the forecast period;
(d) the number of children expected to leave the area over this period.
To these factors we have to add the further complication (e) that
not all the children will go to the State schools in the area, a proportion
in England, for example, will go to Roman Catholic and other private
schools. The situation is summarized in Figure 14.2. To obtain the final
forecast, it is thus necessary to look at the factors and see how they
may each be forecast or measured. Let us look at the factors in turn:
(a) The number of children who are already in schools in the area can
be used to form the basis of the forecast. For example, if the
number of sever-year-old children at the start of one school year is
220 in the area, then the number of eight-year-olds at the start of
the following year will be 220 plus the number of children arriving
less the number of children leaving. This, however, will not work
for the classes of children who start school for the first time during
the period being forecast.
(b) The forecasting of these last numbers is based on the number of
births in the area. If children start school at five, then the number
of births in the area in past years will provide the basic information
for forecasting the number of new entrants to the schools for the
217
Figure 14.2 A structure for forecasting child populations
next 4 or 5 years. This would probably be 4 years, as a year will

usually be lost in getting the accurate information. If we wish to
forecast further into the future than this, we will have to forecast
future births in the area. This can be done at the simplest level by
taking the series of past births in the area and forecasting from this,
using the methods developed for constant mean model, trending
models or whatever seems appropriate. A more complicated
approach would be to seek to form a model based on the factors
thought to influence the birth rate. The simplest way to do this is
to try a simple linear regression model. In one study some fifteen
factors were listed as likely to influence the number of births.
These included such things as the area population, number of new
houses built in the year, previous birth rates and the number of
women in various age groups. The final regression equation used
was
Number of births in year N = 0.006 (population)

+ 0.055 (new houses completed in
year N)
+ 0.515 (births in year N— 1)
+ 0.227 (Roman Catholic baptisms in
year N — 2)
Clearly, such an equation is an empirical attempt; to select the main
factors in the situation and is subject to the various cautions given in
Chapter 6. However, in the above situation it ‘explained’ 80 per
cent of the variation in past data.
(c) In attempting to estimate the number of children arriving in and
leaving the area a major factor is the construction of new housing
and the destruction of old housing. This type of information is
usually available as part of local planning information. The problem
is that of relating the new and old housing types to the number of
children of the various age groups involved. The procedure here is
\
218
to classify the houses according to type, e.g. by the number of

bedrooms and cost range. A survey of new and old properties of
the various types is then carried out to obtain information on the
number and ages of the children in each type. This enables the
housing plans to be converted to forecast the number of incoming
and outgoing children. In one study the above procedure was
elaborated by allowing for the fact that supply and demand
relationships will affect the pattern of occupancy of new housing.
In dealing with the number and age structure of those leaving the
area and those moving internally, information can be obtained via
surveys and also from the records of schools in the area over past
years.
(d) In a number of surveys regression methods have been used to
forecast future numbers of Roman Catholic baptisms. As an
example, denoting the number of baptisms in year N by BN and
the number of births by bN, then one model used was
BN ~ 0.20 bjy + 0.230 Byj _ j +0.32 Bjy _ 2
In estimating the total child population in Roman Catholic schools

in one area the above relation was used for forecasting baptisms of
children in the area. However, numbers of children came to local
Roman Catholic schools from other areas around. Rather than seek
the detailed information required to produce a regression type
model for each of these areas, exponential smoothing was used on
the number of such children attending in past years as obtained
from school records. Thus, though a better method was available
for forecasting, considerations of economy led to a less sophisti-
cated method being used. The loss in quality of forecasts was here
not so important, as the factor concerned was only one of many. A
forecast of the number of children going to private schools can be
obtained by studying past numbers and proportions attending such
schools and using standard forecast methods on the sequences
obtained.
The structure of the situation, as indicated in Figure 14.2, now
enables all the various statistics and forecasts to be brought
together and welded into a total model of the situation which
provides the final forecasts.
References
Albert, A. E. and Gardner, L. A. (1966). Stochastic Approximation and Non-Linear
Regression. MIT Press, Cambridge, Mass.
Bamber, D. J. (1969). A versatile family of forecasting systems. Op. Res. Quarterly,
20, 111-121.
Berrisford, H. G. (1965). Relation between gas demand and temperature. Op. Res.
Quarterly, 16 229—246.
219
Brenner, J. L., D’Esopo, D. A. and Fowler, A. G. (1968). Difference equations in

forecasting formulae. Man. Sci., 15(3), 141—159.
Chen, G. Y. C. and Winters, P. R. (1966). Forecasting peak demand for an electric
utility with a hybrid exponential model. Man. Sci, 12(3), 531—537.
Chow, W. M. (1965). Adaptive control of the exponential smoothing constant. J. of
Ind. Eng., 16,314-317.
Eilon, S. and Elmaleh, J. (1970). Adaptive limits in inventory control. Man. Sci.,
16(B), 533-548.
Harrison, P. J. and Stevens, C. F. (1971). A Bayesian approach to short term
forecasting. Op. Res. Quarterly, 22, 341—362.
Kalman, R. E. and Bucy, R. S. (1961). New Results in linear filtering and prediction
Theory. J. of Basic Eng., 83D, 95—108.
Montgomery, D. C. (1970). Adaptive control of exponential smoothing parameters
by evolutionary operation. AIIE Trans., 2, 268—269.
Morrison, N. (1969). Introduction to Sequential Smoothing and Prediction.
Newbold, P. and Granger, C. W. J. (1974). Experience with forecasting univariate
time series and the combination of forecasts. J. Roy. Statist. Soc. A, 137,
131-146.
Rao, A. G. and Shapiro, A. (1970). Adaptive smoothing using evolutionary spectra.
Man. Sci., 17,208-218.
Roberts, S. D. and Reed, R. (1969). The development of a self-adaptive forecasting
technique. AIIE Trans., 1, 314—322.
Roberts, S. D. and Whybark, D. C. (1973). Adaptive forecasting techniques. J. of
Ind. Eng..
Robinson, E. (1967). Multichannel Time Series Analysis with Digital Computer
Programs. Holden-Day Inc., San-Francisco.
Trigg, D. W. (1964). Monitoring a forecasting system. Op. Res. Quarterly, 15,
271-274.
Trigg, D. W. and Leach, A. G. (1967). Exponential smoothing with an adaptive
response rate. Op. Res. Quarterly, 18, 53—59.
Van Dobbin de Bruyn, C. S. (1964). Prediction by Progressive correction. JRSS (B),
26,113-122.
Ward, D. H. (1963). Comparison of different systems of exponentially weighted
prediction. The Statistician, 13,173—185.
Wheelwright, S. C. and Makridakis, S. (1973). Forecasting Methods for Management.
John Wiley and Sons, New York.
Whittle, P. (1965). Prediction and Regulation by Linear Least Squares. EUP,
London.
Wiener, N. (1949). Extrapolation, Interpolation and Smoothing of Stationary Time
Series. John Wiley and Sons, New York.
Yaglom, A. N. (1962). Theory of Stationary Random Functions. Prentice-Hall.
Englewood Cliffs. New Jersey.
Chapter 15
The analysis and comparison of methods
15.1 Introduction
It would make life much easier if we could say that the forecasting
formula on page X is the best to use and all the others are of historical
interest. Unfortunately we cannot. One might say that surely knowing
the model the theoretician can derive the best forecasting formula. It is
true that he can, theoretically, derive the best forecasting formula given
some criterion, but in practice there are many snags with this. The
model we use will probably not be exactly right and even if it was
good enough for the past it may not be too good now. The model may
only describe certain aspects of the actual situation; other aspects will
shows up in the forecasting errors. Even if all is well we need to have a
forecasting formula that responds well to minor changes that might
occur in the structure of the data. In general, then, we can rarely
choose, for a practical situation, the appropriate method without a
great deal of analysis of the data and a comparison of various possible
methods. The aim of this chapter is to discuss ways of carrying out this
analysis and comparison.
It is possible to approach the comparison of methods in a purely
mechanical fashion; we simply calculate the mean square error for a set
of methods and/or a set of values of the forecasting parameters and
choose the best. However, for effective forecasting the objectives are to
seek the best methods and to understand their properties. It might be,
for example, that a more careful study of the results of the trials would
have led not to choosing that method with the smallest mean square
error but to developing an entirely different method. At the heart of
our study is the forecast error. Most, though not all, of the methods
developed in this chapter will be aimed at analysing the forecast errors
to see if they show any structure of their own. If the forecast errors do
show a clear structure then, as we will see in Chapter 17, we may use
this knowledge to obtain better forecasts or make better use of the
forecasts we have.
221
One or two general observations should be made in relation to the

use of errors in this chapter. Firstly, one must obviously study the
errors for the type of lead time of practical interest. There is little point
and possible dangers in studying one-month-ahead errors if the lead
time in practice is going to be three months. If there are several lead
times of interest they should all be analysed, though if this is not
possible some appropriate middle value should be used. Secondly, in
plotting graphs and calculating statistics it should be remembered that
the first few forecasts are very unlikely to be as good as the later ones.
These should therefore be discarded or somehow weighted before
carrying out the error analysis. Thirdly, if forecasts depend on estimates
of model parameters and/or choices of forecasting parameters made
from a given set of data, then it is unwise to carry out the forecast
analysis on that same set of data. Having said this one must admit that
frequently data is so scarce that this has to be done. The danger, of
course, is that the parameters having been chosen using the data lead to
a fitted model or forecasting formula that fits very snugly to that
particular set of data; however, when it is applied to future data the fit
will probably not be so good and the conclusions of the forecast
analysis may turn out to be wrong. With sufficient past data one can
avoid this problem by choosing the parameters on the basis of the first
half of the data and carrying out the forecast analysis on the second
half. If there is not much data, some improvement can be obtained by
carrying this out as stated but then interchanging the two halves of the
data and repeating the process. This process is called crossvalidation. An
alternative procedure is based on the fact that many models are
essentially unaltered in form by reversing the time axis. In such a case
we might choose the best forecasting parameter by using the data as
though xt were the first observation, xt_ x the second, and so on. The
forecast analysis would then be carried out using the data in the correct
time sequence. When comparing different methods, there is something
to be said for comparing the best possible results obtainable for each
method. These we will get by using the same data for the fitting
exercise as for the comparison. As long as it is realized that the results
will be better than we could expect in practice, then we are on
reasonably safe ground.
It will be assumed in the following sections that we have decided on
an appropriate model and hence have some reasonable methods to try
out. If our choice was poor the results of the analyses to be described
should tell us of the fact and guide us towards more appropriate
choices.
15.2 Forecast analysis

The human eye has a great ability at detecting pattern, admittedly
sometimes erroneously. However, as we are looking for structure, for
\
222
pattern, in our data and in our forecast errors, we would be most

unwise to ignore this ability. The aim of this section is to list a number
of useful ways in which graphical and statistical analysis can help in
forecasting. We will assume that for the method of forecasting under
investigation we have a set of forecasts together with the observed
values that occurred and the consequent errors, denoted by (xiy xu ez),
i = 1, . . . , t. The main types of procedure of practical use are as
follows.
15.2.1 Plots of forecasts and actual observations (xz, xt)
(a) Double plot against time
In Figure 15.1 we have simply plotted forecasts and observations on

the same graph. This gives an immediate impression of how well the
forecasts are following the observed data.
Time
Figure 15.1 Plot of data (•) and forecast (+) superimposed
(b) Forecast—observation diagram
In Figures 15.2(a) the forecasts are plotted against the observations. If

all the forecasts were perfect, all points would lie on the 45° line shown
in Figure 15.2(a), termed the line of perfect forecasts. In the figure
there is a tendency for the points to lie below this line, thus indicating a
bias in the forecasts causing them to be larger on average than the
observed values.
223
Forecast a decrease Underestimate ✓

when an increase an increase /
actually occurs
/ Overestimate
/ an increase
Overestimate / A
z,
a decrease /
Forecast an increase
/ Underestimate when a decrease
/ a decrease actually occurs
(c) Forecast-observation diagram for change
Figure 15.2 Forecast—observation diagrams
(c) Centred forecast—observation diagram
In Figure 15.2(b) we have moved the axes of the forecast-

observation diagram so that they are somewhere near the middle of the
data. The numerical value at the intersection point must be the same
for both x and x, but its exact value does not matter too much as long
as there is a reasonable spread of data to either side of the axes. Again
we have a line of perfect forecasts at 45° to the axes. The actual points
shown on Figure 15.2(b) show a tendency to fall below the line of
perfect forecasts in the upper right-hand quadrant and an opposite
tendency in the lower left-hand quadrant. Thus for large observations
there is a tendency to overforecast and for the smaller observations
there is a tendency to underforecast. An effect of this would be to give
\
224
nearly unbiased forecasts with a larger variance than the original

observations.
The nature of the forecast—observation diagram suggests two
numerical measures of forecast quality. Firstly, we can say that we
would like the slope of the regression line of x on x to be close to that
of the line of perfect forecasts, i.e. a slope of 1.0 and an angle of 45°.
The formula for this slope, from the ordinary theory of regression, is
— x)(x; — x)
Slope = m =
X(Xi-I)2
where x and oc are the averages of xt and xif respectively. The other
measure is the correlation between x and x\ this is given by
rlx 2(x,-x)(Xi-x)
' ’V[2(^-x)22(x,-i)2]
If we had perfect forecasts, x{ = xi9 this would give r(x,x) = 1. Thus a
large value of r is an indication of a good forecast. Notice the relation
between m and r, namely
where sx and s% are the standard deviations of x and x, respectively.

The line of perfect forecasts would thus correspond to forecasts for
which m = 1, r = 1 and sx = s%.
Clearly the values of r and m are of value for studying the quality of
our forecasting methods for all our data. It often happens that local
variations occur in the underlying situation. This has the consequence
that the values of r and m vary locally. We may estimate such local
values by taking moving sets of data for the calculations, as in moving
averages. Alternatively, we can introduce a weighting factor in the
summations for which the weights die away in both directions. This is
most easily done by using weights that correspond to exponential
smoothing. This is illustrated in Appendix C. Figure 15.3 shows local
correlation, r(xt9xt), derived in this fashion, plotted against t.
(d) Forecast—observation diagrams for change

In some situations one is interested in a method’s ability to forecast
increases or decreases in a variable, e.g. whether sales are going to go up
or down next month. Thus we are interested in changes from the
present sales xt. Thus let
sy /y* /y»
1 Xf
denote the observed change in sales and

A A
sy /y /y
*W+1 Xf
225
0 95-
Time
Figure 15.3 Plot of local correlation
denote the forecast change in sales. The quantities zt and zt can now be
plotted on a forecast—observation diagram as in (c) above. The
interpretation of the points obtained are as indicated in Figure 15.2(c).
If points lie in the upper left or lower right quadrants, the inference is
that a ‘turning point’ error has been made; z and z have opposite signs.
In the other two quadrants the errors are in terms of magnitude of
change but not of direction. A detailed discussion of this type of plot,
together with many examples, is given by Theil (1958).
15.2.2 Analysis based on forecast errors

The plots and statistics here are all based on studying the forecast
errors. In the discussion we will use the ordinary error et - xt — xt. It
is, however, sometimes useful to use the percentage error 100 et/xt in
place of et in the analyses.
(a) Tabulating forecast errors

The most obvious analysis of the errors is to simply look at the errors
themselves. Table 15.1(a) shows a small part of such a detailed analysis.
The errors given are those made in forecasting the observation made at
the given time from h time units previously, h = 0 corresponding to
simple smoothing.
Table 15.1 Tabulating forecast errors
(a) Part of a list of errors
Lead time h
Time 0 1 2 3 4
20 -0.004 0.066 0.236 0.407 0.377

21 0.016 0.170 0.324 0.278 0.232
22 0.042 0.154 0.065 -0.024 0.088
23 0.038 -0.089 -0.216 -0.143 -0.070
24 -0.022 -0.127 -0.032 0.063 0.258
(b) A set of standardized errors
Standardized 0.81 1.46 0.84 -0.69 -0.45 0.10 -0.23 -1.92

errors
Ordered -1.92 -0.69 -0.45 -0.23 0.10 0.81 0.84 1.46

standardized
errors
Another useful device is to standardize the errors, i.e. subtract e and

divide by their standard deviation se. If the data are approximately
normal, 95 per cent of these should lie between —2 and +2. Anything
much greater than 3 can be regarded as an outlier. Table 15.1(b) shows
a set of such standardized errors in their order of occurrence and in
their order of magnitude.
(b) Plot of error against time

In Figure 15.4(a) the simplest plot of et against time is shown. This
gives the clearest indication of the nature of the forecast errors. It will
indicate their magnitude, roughly how independent adjacent errors are,
the existence of trends or other systematic changes in the errors and the
occurrence of any particularly large errors.
(c) Histogram of forecast errors

The first question about the statistical properties of errors concerns
their distribution. The histogram or bar chart such as that in
Figure 15.4(b) provides one of the easiest presentations of the sample
distribution. The area under each bar is equal to the frequency of errors
occurring in the corresponding range of values of e. If equal intervals of
e are used, as in the example (b), the vertical height can be used to
represent frequency.
227
Frequency
ro <3- loinrooj — o
ro <\j — i — CM ro
1
l I
Lower class limit
Figure 15.4 Plots of forecast errors
(d) Normal plots
Suppose the histogram looks reasonably symmetrical and it is desired

to construct some sort of limits to forecasts based on assuming
normally distributed errors. In this situation it is useful to carry out a
further check on the normality of the error distribution. This can be
done with the use of normal plots, such as that illustrated in
Figure 15.5(a). This may make use of commercially available graph
paper. The axes of this paper are so designed that plots of the data are
approximately straight lines if the data come from a normal population;
otherwise it reveals some form of twisting. Figure 15.5(b) indicates how
to interpret the standard normal probability plot.
\
228
o>
~D
<V
M
o
“O
c
o
CO
(a) Normal plot of errors
(b) Interpretation of deviation from straight line
Figure 15.5 Normal plots
(e) Errors against forecasts

This plot, illustrated in Figure 15.6, gives an indication as to whether
there is any relation between the size and sign of the errors and the
magnitude of the forecasts. If we regard the forecast as a rough estimate
of the underlying mean, then this plot gives an indication as to whether
the magnitude or sign of the forecast errors are effected by the
magnitude of the underlying mean. An immediate application of this
occurs in the distinction between the additive and the multiplicative
occurrence of the irregular component in the seasonal model. In the
additive form the errors are not influenced by the size of the seasonal.
If, however, we have a situation which is multiplicative, then the
229
• 0
0
•
• •
•
• •
• •
•
o
•
<D • • •
~o
• •
a;
.S
"O
0 - •• •
•
• 00
o • •
X)c •
o • •
CO •
• • 00
• •
•
0
•
•
•
«
-2
•
Forecast
Figure 15.6 Plot of error against forecast
forecasting errors will tend to be greater when the seasonal peaks occur.
This will show up as a greater spread in the errors for large x.
(f) The distributions of x and x

The plots that we have used to look at forecast errors can obviously
be used to look at the distributions of the observations and of the
forecasts. If the means of x and x are following some deterministic
curve, a trend or a seasonal for example, such plots will not be very
meaningful. However, even if we do have such a situation, it is still
informative to study the conditional distribution of xt+h, given xt+h.
Thus we are concerned with finding the distribution of xt+h knowing
the value of our forecast xt+h, e.g. with the distribution of the future
sales relative to our current forecast of their value. We denote this
distribution by f(x j x). Referring to the forecast—observation diagram
(see Figures 15.2 and 15.7), this distribution is the distribution of the
observation x lying in the line of fixed oc.
Without a vast amount of data or an assumption of normality this is
very difficult to obtain directly. We can, however, proceed in a
different fashion. It can be argued both theoretically and intuitively
that the forecast errors in many situations should be independent of the
magnitudes of forecasts. Plots such as Figure 15.6 give an immediate
visual check on this. Our distributional studies of the forecast errors in
(c) and (d) should give us some idea of the error distribution f(e).
Because of the independence assumption this is identical to the
conditional distribution of the error, given x,f(e I x).
\
230
(b)
Figure 15.7 Conditional distribution of xt+ ^

given xt+h (a) (b)
Writing e = x — x we now know the distribution
f(x li)
Thus the distribution of the future xt+h is obtained by replacing et+h

by xt + h — xt + h, xt+h being a known forecast, in the distribution of
et+h. This is seen with reference to Figure 15.7(b). Provided that the
distribution of et+h does not depend on xt+h the distribution ofxt+h
about xt+h, on the line of perfect forecasts, is simply the distribution
of et+h . The simplest example is the case where et+h is normally
distributed, N(0,ol). It follows that xt+h is also normally distributed,
N(5ct+h, aj).
(g) Comparison of errors from different methods

Instead of comparing just the mean square errors of different
methods it is often useful to compare the individual errors. This is
done, as in Figure 15.8, by simply plotting the errors for both methods
against time or on a scatter diagram.
231
20
Method I o
Method 2 x
o o
x
10 X
X O
X
o X
X o
X X X
GO
o O
w_
0 -f
LLI I 2 3 4 5 6 7 8 9 10 II 12 13 14 15 Day number
x
Mon Mon Mon O
O
o x
-10 o
x
o
X
o
-20
X
x
o o
(a ) Direct plot against time
Method 2
error
Figure 15.8 Comparison of errors from different methods

\
232
Figure 15.9 Autoregression of errors
(h) Autoregression of errors

If the sequence of errors ex, e2, . . . , et is put in the form of t — 1
pairs, (eue2), (e2, e3),. . . , (et_2, e(_,), (et_uet), the correspon-
ding points can be plotted on a scatter diagram as in Figure 15.9. If
successive errors are independent of each other, the plot obtained will
show a completely random scatter. If the points look like those
illustrated, there is an indication of a relationship between successive
errors, a relationship that may be used to improve future forecasts. The
general term for such a relationship is an autoregression.
(i) Autocorrelation of errors

As a measure of the extent of the linear relationship between
successive errors in the scatter diagram of Figure 15.9, we can use the
correlation coefficient between the pairs ( eh ei+l), i = 1,. . . , t — 1.
We call this the autocorrelation coefficient of lag 1, rx. If we had used
pairs two apart, (ef, e,-+2)» * = 1, . • . , t — 2, this would give the
autocorrelation coefficient of lag 2. Thus in general we can calculate
^k ~ î+ k )
If we have a good forecasting method, it is reasonable to require that
the forecast errors are independent of each other. This implies that
expected values of rx, r2, . . . are all zero. To check this we calculate rL,
r2, . . . and plot these as shown in Figure 15.10. Such a plot is called a
233
0 5
« 0-4
o 03
o 0 2
O) 0 1
12 13
o o
11
1 I 1__L i—i— Lag
8 9 10 II
o'OI
D -0 2
<-0 3
-0 4
-0 5
(a) A good result-small autocorrelations
Lag
(b) A result suggesting possibility of improving

forecasts
Figure 15.10 A correlogram
correlogram. If any of the r values are significantly different from zero,

the forecasting method can probably be improved. One suitable
method is discussed in Chapter 17. A particular application of this plot
is that of testing the success of a seasonal forecasting model. If we have
failed to adequately model the seasonal components, the autocorrela-
tion of lag corresponding to the length of the season will tend to be
large.
15.2.3 Root mean square error analysis

The root mean square error is the most commonly used forecasting
criterion. Thus several approaches to forecast analysis depend on a
study of the root mean square error (RMSE).
234
Table 15.2 Comparison of root mean

square errors
Lead Method
time I II III
0 0.0361 0.0358 0.0286

1 0.1451 0.1449 0.1439
2 0.2144 0.2115 0.2101
3 0.2809 0.2817 0.2816
4 0.3325 0.3342 0.3347
5 0.3961 0.3981 0.3982
6 0.5009 0.5021 0.5016
7 0.6177 0.6183 0.6153
8 0.7268 0.7273 0.7237
9 0.7681 0.7696 0.7602
10 0.8822 0.8855 0.8723
Number of errors used equals 37 minus

the lead time.
(a) Tabulation
The most obvious step is to tabulate the root mean square error or
the appropriate criteria for the problem. Table 15.2 compares the root
mean square error for three methods for different lead times. Notice
how the root mean square error increases with lead time. A natural step
following this tabulation is to seek to model this increase relating the
root mean square error to the lead time. In Table 15.2 a model giving a
linear relation would not be far out.
(b) Plot of local root mean square error
As mentioned in section 15.2.1(c) and illustrated in Figure 15.3, we

can readily calculate a local value of any statistic using the weighting
technique described in Appendix C. Thus, instead of relying on a single
value of the root mean square error for all our data, we can calculate its
local value and plot this quantity against time. Figure 15.11 shows such
a plot. We see in this example a clear indication that our forecasting
method has been getting increasingly inaccurate.
(c) Plots of local bias and standard deviation

We have noted before that the mean square error can be broken into
contributions from the bias and from the error variance. Thus, in
theoretical, population, terms
E{e2 ) = E(e)2 + Var(e)

235
i±j
CO
s
cc
o
o
o
130 —
120
Time
Figure 15.11 Plot of the local root mean square error
Having looked at the local root mean square error, it is natural to see
how its fluctuations are related to the separate local fluctuation in the
bias, E(e) and the standard deviation of error, J Var(e). Figure 15.12(a)
and (b) show such local plots for the same data as that used in
Figure 15.11. It is clear from these that the fluctuation in the root
mean square error is produced almost totally by that in the standard
deviation. The bias is very small and its local fluctuations have little
effect on the root mean square error. Had we been in a different
situation and had large local biases occurring, then we could improve
our forecasts by making some adjustment to remove the bias. We will
examine this possibility in Chapter 17.
(d) TheU's coefficients

Having split the population mean square error into components, we
can in fact go a stage further and split the error variance into two
components. Thus
Var(e) = Var(x — x)
= al + a\ —2 axGip(x,x)
using a standard statistical result. Here ox and a* are the population

standard deviations for observations and forecasts and p{x,x) is the
236
Figure 15.12 Plots of local bias and standard deviation

237
population correlation coefficient. Replacing these by our estimates sx,

and r(x,x)9 and noting that the bias from the data is just the average
error, e, we have
MSE = e2 + s2x + sj - 2sxs$r
A little algebra will show that we can rewrite this as
MSE = e2 +s? (l —rM

X 1
= bias + regression + correlation

term term term
These three terms have a clear interpretation. The bias, as we have seen,
indicates under- or overforecasting. The regression term, from the
results of section 15.2.1(c), measures how much the regression line of x
on x, in the forecast—observation diagram, differs in angle from the 45°
line of perfect forecasts. The regression term is, in fact, s| (1 — m), m
being the slope of the regression line. The correlation term depends on
r, the correlation r(x,x) between observations and forecasts. If we
divide all through our expression by the mean square error, the
left-hand side becomes one and the positive fractions on the left-hand
side will indicate the proportion of the mean square error that is due to
the bias, regression and correlation aspects of the forecasts. The three
terms in this form are called Theil’s coefficients (see Theil, 1958 and
1965, and also Granger and Newbold, 1973). As the three coefficients
add to one, they can be plotted on triangular graph paper as in
Figure 15.13. The mean square error for the forecasts leading to points
Aj and A2 is due mainly to poor correlation between forecasts and
observation. For B2 the mean square error is due largely to a bias in the
forecasting. For the mean square error is due to both bias and poor
correlation. For and C2 all three factors contribute to the mean
square error. Note that these interpretations depend on the easily
verified assumption that si and s\ are roughly the same.
15.3 Choosing forecasting parameters

We have so far discussed error analysis from a general investigational
point of view. We turn now to the much more specific aim of the
choice of the numerical values of forecasting parameters. It is assumed
here that the method has been decided upon and the only task left is to
choose these parameter values. The essence of the operation is to decide
on the forecasting criterion to be used and then, in some fashion, to
find the value of the forecasting parameter for which this achieves its
best value. In the great majority of practical analyses of this type the
mean square error is used as the criterion, though a weighted mean
\
238
i-o
Figure 15.13 Plot of Theil’s coefficients
square error may be appropriate where there is a long sequence of past

data. In most of the forecasting formulae we have discussed there have
been only one or two forecasting parameters. It is also clear that from
the nature of these parameters we do not require very accurate
specifications of them. For example, with most applications of
exponential smoothing a can be chosen to within 0.05 quite satis-
factorily. Hence there is usually no need to conduct any sophisticated
search to find optimum values. The simplest approach is just to use past
data to evaluate the mean square error at a lattice of points of possible
interest and plot the result.
Table 15.3 gives a number of illustrative examples. From these a
number of general points can be made:
(a) Where the forecasting parameter is the parameter a from dis-
counted least squares then, if a0 is the optimum,
(i) a0 = 1 corresponds to ordinary least squares and the model is
being used as a global model.
(ii) for stochastic and regression models values of a0 close to one
(a0 — 0.9) are the norm.
239
Table 15.3 Values of mean square error for various discounting

factors a
a
Data 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1
Set 3 46 42 39 36 33 32 31 30 31 32
Set 2 86 85 87 92 97 103 — — — —
Set 1 39 40 42 45 49 55 — — — —
(iii) if is small, e.g. a0 = 0.5, or much greater than one, e.g.

G0
a0 = 1.3, then this is taken to indicate that the model being
used is inappropriate.
(iv) the curves of the criterion function may be fairly flat neara0.
This, together with the fact that a0 is a sample quantity with
its own random variation, gives the forecaster some leeway in
fixing the value of a to use in practice.
(b) Where there are more than one parameter, a more elaborate lattice
is needed, as, for example, with Holt’s method, see Table 15.4(a).
If several parameters are to be optimized for a large number of sets
of data, more efficient ‘hill climbing’ methods should be used.
(c) The choice of model out of a class of models, e.g. of p for an
Table 15.4 Choosing two forecasting parameters
(a) Parameters a and b in Holt’s method
MSE Mean parameter a

tabulated 0.5 0.6 0.7 0.8 0.9 0.99
0.5 368 352 345 351 359 370

0.6 341 326 324 328 331 358
Slope 0.7 326 313 305 316 319 337
parameter b 0.8 314 302 294 299 308 320
0.9 305 293 285 289 294 390
0.99 315 305 296 302 312 322
(b) Discounting parameter a, order p, of autoregressive model
Order p
MSE
tabulated 1 2 3 4 5 6
1.0 925 916 907 891 853 848

Discounting
0.98 890 880 869 867 814 824
parameter a
0.96 901 884 879 878 823 847
\
240
autoregressive model AR(p), is structurely equivalent to choosing a

forecasting parameter, as is illustrated by Table 15.4(b).
15.4 Comparison of methods

We now turn briefly to the comparison of methods. It might seem odd
to deal with this after discussing the choice of forecasting parameters,
however a little thought shows that if we are to compare methods we
must first choose the best parameter values for those methods. We thus
compare the best results that each method can achieve. The comparison
of three methods in Table 15.4, for example, in fact corresponded to a
comparison of double exponential smoothing, discounted least squares
trend fitting and extrapolation and Holt’s method. In the first two cases
the single parameter was chosen to give the smallest mean square error
for a particular lead time and in the latter case the two parameters were
chosen in this way. Such choices of the best method use one given
criterion, but that does not mean to say that this criterion must be the
only quantity used in the comparison. Indeed, if we wish to compare
methods for a given type of data, it would be unwise to compare them
only by error analysis on past data; some theoretical and simulation
comparison should also be carried out to study the behaviour of the
methods under varying circumstances. As we indicated in Chapter 3, it
is very important to ask questions about the stability and robustness of
one’s methods; this is particularly important if the methods are to be
used on a regular basis, say as part of a computer program for stock
control. If one has a good set of past data some useful information on
these aspects can be gained by adding, for example, step-changes or
trends to the real data, forecasting these ‘adulterated’ data and then
carrying out an error analysis. If one is using a simple linear forecasting
formulae, which depend simply on some form of weighted sum of the
observations, the new forecast equals the old forecast plus the forecast
of the step or trend. This exercise would not then provide any more
information than that obtained by a general study of the method,
without reference to ones own data. However, many methods,
particularly those involving seasonal aspects, are not linear and useful
information about the methods being studied can be obtained by the
use of ‘adulterated’ data. If more extensive studies are required, then it
is best to simulate data from a range of situations that might
foreseeably occur in the practical field being studied. These data would
then be used in the various methods of forecasting and error analyses
carried out.
A number of papers have appeared that compare both theoretically
and practically some of the methods discussed in this book (e.g. Davis
and Huitson, 1967; Godolphin and Harrison, 1975; Granger and
Newbold, 1973; Groff, 1973; Marchant and Hockley, 1970; McKenzie,
241
1974; Newbold and Granger, 1974; Wagle, Rappoport and Downes,

1968). Rather than discuss these papers, the points made about the
various methods have simply been included in the relevant sections of
the book. However, the point that does emerge from this and other
work is that there is certainly no best method for all situations and that
what is best by one criterion may not be by another. The problem is
that there are many features of a method to consider and not just its
mean square error properties. One method will produce a forecast of
reasonable properties as soon as three or four observations are available;
another method will require sixty observations before it can be really
started. One method will require only storage of the last observation for
each series being forecast; another will require the last thirty observa-
tions. One method can be implemented and understood by a clerk;
another will require a medium sized computer and a person with a
mathematics degree to understand it. One method will perform well at
short lead times but badly at long ones. One method will not be much
affected by a couple of wild values; another will be badly out for a long
period after the trouble. One method will involve the use of several sets
of information that have value on their own and will produce useful
ancillary information; other methods will just produce the forecast.
Clearly, with so many aspects to consider, each forecaster must
choose his own methods to fit his own situation. A brief survey of
individual practice shows that this is exactly what happens. There are
nearly as many methods and approaches as there are forecasters. This
sometimes produces a feeling of a ‘Tower of Babel’, but the above
comments should indicate that it is not entirely due to human
cussedness.
15.5 Prediction intervals

Suppose we are in the position of having a satisfactory forecasting
method and a forecast of xt+h obtained from it. What can we say about
xt+h? In particular, can we give an interval, a prediction interval
(xt+h e> xt+hfU) in which xt+h has a high, known, probability of
falling? There are two possible approaches to solving this problem: one
theoretical, based on the model used, and one empirical, based on the
observed behaviour of past forecast errors. Let us consider these in
turn.
(a) Intervals derived from the forecasting model

We first came across the idea of confidence limits as a means of
getting a prediction interval in section 4.1. In essence the discussion
given there can be applied in any situation where we can use the model
to find the variance of the forecast error and where that error can
\
242
reasonably be taken to be normally distributed. In past chapters we

have looked at a range of such formulae from
t+ 1
Var(e) =
t
for the constant mean model, with the average used as a forecast, to
Var(e) = Var(6)x2 + a2 + Var(xX Var(b) + (32 }
when forecasting with a regression model; in many of these cases the

normality of e in the model leads directly to normally distributed
errors. In general, given Var(e) = a2 as the known variance of e, we can
write statements such as
Prob( — 1.96 oe < e < 1.96 oe) = 0.95
and hence to
Prob (x — 1.96 oe < x < x + 1.96 oe) = 0.95
Thus we have upper and lower values that on 95 per cent of occasions,
when using the method, will enclose the value of x. In using the model
we usually find a2 as a function of o2, the variance of the e values in
the model. This variance has to be measured in terms of the observed
variance when fitting the model. In this case a2 is replaced by its
estimate, s2 , and 1.96 has to be replaced by the value from the tables of
the f-distribution with degrees of freedom as required by the estimate
s2. Where the errors are not normally distributed the problem is more
complex.
Often in standard situations using global models standard computer
programs derive these intervals. The basic problem here is that these
calculations are usually based on assumptions of global rather than local
models. Thus any deviation of the actual structure from the model will
mean that the calculated prediction interval will not lead to the correct
proportions of observations inside the interval. For example, in a
situation described by Williams and Goodman (1971) only about 80 per
cent of observations lay inside the 90 per cent prediction intervals
calculated using an assumed, regression, model.
(b) Intervals derived from past forecast errors

The second approach here is to study the distribution of past
forecast errors and estimate o2 by the ordinary estimate, s2, of
variance. If the errors are approximately normal with zero mean, then
we will have for the 95 per cent prediction intervals, using n errors,
Prob(x — t„ _ , s < X < x + tn _ !

>9 5 % e >9 5 %se) = 0.95
243
Though there are problems if the errors are not normally distributed,
the basic approach is the same. In Williams and Goodman (1971),
referred to above, the magnitudes of the errors, I e I, were found to give
a good fit to a gamma distribution. This distribution was fitted to the
set of errors and prediction intervals obtained. The observed propor-
tions of forecasts, other than those used to fit the model, that fell
within the intervals was satisfactorily close to the probabilities as fixed
from the gamma distribution. An alternative approach, that makes no
assumptions about the distribution, is based on simply ordering the
errors. Suppose, for example, we were concerned only with estimating a
value for e that had only a probability of 0.1 of being exceeded. In this
case the largest of ten observed errors, or the second largest of twenty
observed errors, etc., would provide a rough estimate based on the
observed errors alone. To use estimates of this form we need to be
fairly confident that the data used to obtain these estimates come from
a stable error distribution and do not contain any outliers.
15.6 Sensitivity analysis

Thus far in this chapter we have talked about the study of single
forecasting methods in isolation. Often our forecasts are components of
a complex situation and we may well be using several different
forecasts. In such situations the assessment of the methods and the
study of the influence of forecasting errors on the situation become
more elaborate processes. The term ‘sensitivity analysis’ describes the
sort of exercise the forecaster must carry out in such situations. As an
example of this we return to the example of the use of mixtures of
methods in forecasting school populations given in section 14.6.
Suppose the forecasts being used are:
ht = forecast number of children of given age in region at beginning
of year t
lt = forecast number of leavers during year t
at = forecast number of arrivals during year t
rt - forecast proportion of age in Roman Catholic schools
pt = forecast proportion of age in other private schools
st = forecast number of children in State schools
Then, allowing for only these factors, we have
ht = ht_ i + at_ ! — tt_ i
st = ht{l —Pt—rt)
Clearly, even if we had sufficient past experience to be able to quote
mean square errors for the various forecasts, h, . . . , p, we would have
great difficulty in evaluating the final mean square error for st. This
\
244
Table 15.5 Sensitivity analysis
Line No. "t-1 &t-1 U-1 *t Pt ?t St

1 2,300 100 50 2,350 0.05 0.05 2,115
2 2,500 200 30 2,670 0.05 0.05 2,403
3 2,300 100 50 2,350 0.04 0.04 2,162
4 2,500 200 30 2,670 0.04 0.04 2,456
5 2,300 500 50 2,750 0.05 0.05 2,475
6 2,500 500 30 2,970 0.04 0.04 2,732
could be done with some effort if all the forecast errors were
independent of each other. However, in practice this is unlikely to be
the case. In the above situation there is very little experience of
forecasting and the costs involved are sufficiently high to discourage
much purely experimental work. A natural reaction is to simply ignore
the error analysis side of forecasting and hope that the forecast values
will be reasonably good. There is clearly no hope of giving confidence
intervals for the forecast of st. While this may all be true it is still
nonetheless possible to investigate the forecasting method and say quite
a lot about the behaviour of the forecasts. One line 1 of Table 15.5 we
give the forecast values of all the variables. On line 2 the values of ht_
at_1 and lt_1 have all been set at what in the practicalities of the
situation are thought to be the most extreme values in terms of making
nt large. These are perhaps our guesses at the 95 per cent confidence
limits. The values of pt and rt are left alone. On line 3 we take the line
1 value of ht but let pt and rt take on low values. The sizes of the
schools involved suggest only small drops in p and r are likely and also
puts clear upper values on them. On line 4 we take these low values
with the high value of nt of line 2. By this type of exercise we can begin
to appreciate the effects on the final answer of variations in the
forecasts. If one is willing to make experimental assumptions about the
distributional behaviour of the forecasts, e.g. that ht_ x, at_ x ,\t_ x are
independent with means as on line 1 and standard deviations 100, 50
and 10, respectively, then we can go even further. Clearly, ht will have
mean 2,350 and standard deviation
\/(1002 + 502 + 102) = V(12,600) ~ 112
To find out something of the behaviour of at we must carry out a

simulation. We could, for example, simulate ht from a normal
distribution A(2,350, 12,600) and let pt and rt be normal N(0.05,
0.01), with a correlation of, say, 0.5. On this basis many values of st
could be generated and their mean, standard deviation and general
distribution investigated. The essential part of this type of investigation
is the need for the expert in the particular area to quantify his
245
experience and intuition so that the sensitivity analysis can be realistic

in its comparisons of the various possible futures.
Another aspect of a sensitivity analysis is the possibility of
comparing the results of various possible decisions or assumptions.
Suppose, for example, there is the possibility of a new housing estate in
the area which would generate an additional 400 ‘arrivals’. The effect of
this on the forecast mean situation of line 1 and the extreme situation
of line 4 are shown in lines 5 and 6, respectively. We have taken for our
illustration of this approach a situation which, though complex relative
to the previous situations discussed in this book, is comparatively
straightforward. We can guess the magnitude and signs of the effects of
various changes possible. In more involved situations this is no longer
the case and more elaborate sets of assumptions have to be investigated
in the sensitivity analysis.
References
Davis, R., and Huitson, A. (1967). A sales forecasting comparison. The Statistician,
17, 269-237.
Godolphin, E. J., and Harrison, P. J. (1975). Equivalence theorems for polynomial-
projecting predictors. J. Roy. Statist. Soc. B, 37, 205—215.
Granger, C. W. J. and Newbold, P. (1973). Some comments on the evaluation of
economic forecasts. Applied Economics, 5, 35—47.
Groff, G. K. (1973). Empirical comparison of models for short range forecasting.
Management Science, 20,22—30.
Marchant, L. J., and Hockley, D. J. (1970). A comparison of two forecasting
techniques. The Statistician, 20, 35—64.
McKenzie, E. (1974). A comparison of standard forecasting systems with the
Box—Jenkins approach. The Statistician, 23,107—116.
Newbold, P., and Granger, C. W. J. (1974). Experience with forecasting univariate
time series and the combination of forecasts. J. Roy. Statist. Soc. A, 137,
131-146.
Theil, H. (1958). Economic Forecasts and Policy. North-Holland Publishing Co.,
Amsterdam.
Theil, H. (1965). Applied economic forecasting, North-Holland Publishing Co.,
Amsterdam.
Wagle, B., Rappoport, J. Q. G. H., and Downes, V. A. (1968). A program for
short-term sales forecasting. The Statistician, 18,141—145.
Williams, W. H., and Goodman, M. L. (1971). Constructing empirical confidence
limits for economic forecasts. JASA, 66, 752—755.
Chapter 16
Forecast control
16.1 Quality control in forecasting
We have seen that the essence of statistical forecasting is the

extrapolation into the future of some structure in the past and present
situation. If our assumption that the given structure will continue into
the future is correct, we will get good forecasts but not perfect ones,
since in the real world the structure will always contain some random
element which is unpredictable. The forecast errors in such a situation
will have some statistical distribution which remains constant. We will
sometimes get large errors in this situation, but only with the
probability given by this distribution. A situation which has this type of
stability is said to be in ‘statistical equilibrium’. If we have a stable
structure and any fixed forecasting formula, the forecasting errors will
be in this type of statistical equilibrium. Figure 16.1 showrs a set of
forecast errors from such a situation. An initial set of data yields a
standard deviation of 130, using a zero mean, and a sample mean close
to zero. The histogram of the errors suggests that a normal distribution
would provide a reasonable model for the distribution of the errors and
further study supports this. As there are a large number of observations,
the sample standard deviation can be taken as a good estimate of the
population standard deviation and the mean may reasonably be taken
as zero. From the table of the normal distribution it is seen that 95 per
cent of the observed errors should lie within 1.96 (approximately 2)
standard deviations of the mean. Lines corresponding to these values,
i.e. - 260, are drawn in Figure 16.1. Such lines are called control limits.
As long as the situation remains in statistical equilibrium the errors will
only fall outside these two sets of limits on an average of 5 per cent of
occasions. If, however, a change occurs in the structure, perhaps
suddenly, perhaps gradually, the statistical equilibrium will be lost or a
new equilibrium developed. In either case the statistical structure of the
errors will change. This change may be revealed in many ways, by a
246
247
400
300
Upper control limit

200
100
0
LJ
-100 • •
-200
Lower control limit
-300
-400
Time
Figure 16.1 Control chart for error
change in the mean error or the error standard deviation and a

consequent change in the proportion of observations outside the
control limits. When such a change in the situation occurs the
forecasting method needs to be modified. This modification may take
the form of a change in the parameters of the forecasting formula, a
complete change in the forecasting formula or just a restart of the
method with new initial values.
The use of a graph as a simple visual check on statistical equilibrium,
as in Figure 16.1, has much to recommend it. Unfortunately, in many
forecasting applications the situation is rarely stable long enough for
the calculation of control limits; nor, indeed, is this the most efficient
way of carrying out a control of the forecast errors. For example, if
there is a small change in the mean this will not show very clearly when
only the individual observations are plotted on the chart of Figure 16.1.
It is intuitively better in this case to combine the errors into some
measure of the change in mean. This measure will have known
statistical properties when the situation is in equilibrium. In general,
such measures will be chosen so that for particular changes in structure
there will be clear changes in the measure used. It is clear that for any
given measure, such as the mean error over the available data, certain
changes will show up clearly but others will have little effect. A sudden
change in the mean will in time appear clearly as a change in the mean
error. But a change in structure which takes the form of, say, some type
of added periodic term may not have much effect on the mean error. In
the next section we will consider some measures of error structure
which have proved particularly useful in the practice of forecast
control.
16.2 Tracking signals for bias

The simplest way to keep a watch on the stability of a forecasting
system is to use what has been called a Tracking signal’. This is some
quantity calculated from the forecast errors and such that when the
situation is stable the tracking signal usually remains between specified
limits, but when the situation ceases to be under control the tracking
signal goes outside these control limits. The first suggested tracking
signal was that of Brown (1959) and is based on using the sum of
forecast errors. An alternative to this was suggested by Trigg (1964)
which, though it is based on a very similar formula, has a number of
advantages over the original suggestion of Brown. Trigg’s tracking signal
is based on the use of a smoothed mean of errors. Thus if et represents
the forecast error at time t and et the smoothed mean error at that
time, then
et = (l~b)et + bet_1
where b is the smoothing constant having a value fairly close to one.
Assuming a stable situation in which et has expectation zero and
variance o2 and taking the expectation and variance of both sides of
this equation gives, for large t,
E(et) = 0
and
Var(ef)
1 + b
It is assumed here that the et are approximately uncorrelated. This
assumption usually holds in practice, but it should be checked when
using such a tracking system.
The smoothed error is used to provide a quality control procedure
for the forecast errors. If the errors can be treated as approximately
normally distributed, then 95 per cent of the values of the smoothed
error should lie within approximately — 2og. However, a, and hence a?,
is not known and has to be estimated. Rather than estimate o2 by the
classical root mean square estimate, it is easier in this situation to use an
estimate that involves the minimum of calculation and can easily be
recalculated as each new error is obtained. Brown has suggested the use
of a smoothed version of the mean deviation of errors, which is called
the mean absolute deviation (MAD). Denoting the value of this at time
t by MAD,, then the formula for calculating the MAD is
MAD, = (1 - b) | et \ + b MAD,_ x
249
where | et\ denotes the absolute (signless) value of et. The smoothing
constant is the same as that used for et, for reasons that will become
apparent later. It can be shown that for a normal distribution
E(\et |) =
and Brown (1959) shows that for the normal, exponential, rectangular
and triangular distribution
E (| et|) — 0.8 o — a/1.2
Since E (MADt) = E( \ et\) the quantity 1.2 MADr provides an estimate

of oe. Thus 2 oe control limits would become -2.4 MADf. These
control limits will vary as MAD? varies and will be recalculated
everytime a new observation is obtained. Rather than have variable
limits the smoothed error is divided by the quantity MADf, giving a
quantity which will have fixed control limits. Thus the tracking signal
used is
Tt = et/MADt
To obtain the control limits for Tt we note that, using the 95 per cent
control limits and the relations derived above, the situation is assumed
to be in a state of equilibrium when
—2 Og < et < 2a^

1-b /1-b
— 2 / oe <et <2 / of
1 + 6 1+6
1-6 /I — 6
-2 / 1.2MKDt<et<2 / rr~7 1.2 MAD,
1 + 6 1 + 6
1-6 /I — 6
- 2.4 / < Tt < 2.4
1+6 1 + 6
If T falls outside these limits then there is a suggestion that all is not
well. We do not put it more strongly than this, since 5 per cent of
observations will by chance fall outside these limits even when all is
well. One assumption behind this calculation is that MAD* gives a good
approximation to oe. This is only reasonable when t is large and 6 is
close to one.
A problem associated with the use of this tracking signal is that both
e0 and M AD0, the initial values in the recurrence relations, need
specifying before the data ex, c2 become available. A reasonable value
for e0 would be zero and for MAD0 would be \J2ITI a0, where a0
\
250
would be an estimate of the error variance. Some guide as to the value

to put for 60 can be obtained from knowledge of the forecasting
formula and of the probable data, for example, if the recurrence form
of exponential smoothing is to be used on the data, i.e.
x, = (l —y)x, + yxt_ j
then the variance of xt is
i-7 r2
°l = o;
1+7
assuming independent observations with locally constant mean and
variance o2. Hence the error variance is given by
Var(et) = Var(xf — xt_ j)
1-7 „
= o% + oX
1+7
ol
1+7
By carrying out this analysis we have not gained any information but
have simply replaced the requirement for a guess at MAD0 by a guess at
a2, which is usually more within one’s intuitive grasp of the situation.
If by comparing the variable being forecast with other similar variables
we estimate ox by dx, then the initial value for MAD is
2 ox
MAD0
+
^ 7)}
A better solution to this problem is to defer the commencement of the
scheme until enough actual data are available to provide an estimate of
the mean deviation of errors and to use this value, i.e.
MAD0 =2 | e |/(number of observations)
available
data
We can, however, avoid these starting value problems by using a slightly

different form of tracking signal.
The exponential smoothing recurrence relation which has the form
Zf = (1 — b)zt + bzt_ ,
for some sequence zt, is basically an approximation, when t is large, to
the discounted average
1
t- I A-
2,= 2 Vzt_r 2 V
r- 0 / r=0
251
Using this exact form in the tracking signal gives
t—
/
t- 1
r
2 b e
/
t- t-
2 r
b \e,_r 2 br
/
t- I /*- i
r
2 & e,_r/ 2 6r|e,-r
= Sf(e)/Sf(|e|)
where Sf ( ) represents the exponentially weighted sum, e.g.
t- I
= St(z)= 2 brzt_r
The advantage of this form for the tracking signal is that there is no
requirement for starting values in recurrence relations; T\ depends only
on the data available and not on our estimates. Strangely enough,
though this is the ‘exact’ form it is easier to calculate since quantities
like St(z) have a very simple recurrence form, namely
St(z) = z, + 6Sf_1(z)
The starting value for this recurrence relation is simply S0 (z) = 0 so that
Sx (z) = z j. Obviously if t is large then T't and Tt are equivalent.
Let us examine the properties of Tt and T\ in a little more detail.
Suppose that the process was completely out of control from the start
so that all the errors had the same sign. If the sign was positive there
would be no difference between elf e2, . . . , et and \ ex\, | e2 I, . . . ,
| et |, and hence between et and MAD?, so Tt = 1. If the sign was
negative then et = — I et I, i = 1, . . . , f, and so Tt = —1. As these are the
two extreme possibilities it is clear that T, and equivalently T \ must lie
in the interval (—1, 1). This useful condition is the result of using b as
the smoothing constant for both et and MADf.
In the derivation of the control limits it was pointed out that et was
assumed to be normal and MADf was treated as a constant; this implies
the assumption that T is normal, which cannot be the case with T lying
between limits. The validity of a normal approximation for T must be
checked and it is probably easiest to do this by simulation. Thus we
may generate a long sequence of random variables e having the sort of
property we expect from our actual situation when it is stable. The
sequence of T values calculated from the e values is then examined and
a cumulative frequency polygon constructed. From this we can read off
values of T which will only be exceeded with any required probability.
A discussion of the distribution of T is given in Brown (1963).
16.3 Cumulative sum methods
It is difficult to detect small changes in the mean error using the type of
control chart discussed in section 16.1. One way of improving the
sensitivity is to base the control chart on the averages of several errors
instead of on the individual errors. If we could only guess where a
change in mean had occurred, then the average of all the errors
obtained since the change would be most sensitive in indicating that the
change had occurred. Unfortunately, we do not know when the change
occurred, if it did, so we do not know how many observations to
include in our average. One way of avoiding this problem is to have a
whole battery of control procedures based on the averages of all past
sets of errors, one using just the last error, one using the average of the
last two, one the average of the last three, and so on. Each of these
averages will be the most sensitive available for detecting a sudden
change in the underlying mean error, provided the change in mean
occurred immediately before the earliest error used in that particular
average. If we have say 100 observations, this process effectively
involves the construction of 100 control procedures. This is obviously
impractical but the basic idea does lead to some very useful methods of
control. The first modification required to obtain a practical method is
to replace the average errors by the sums of errors. These are easier to
calculate and are easily related to each other. Thus the quantities used
as the basis for the procedure are
= e
51 t
52 = et + et_ j
53 = et + et_ j + et_2, etc.

which are called cumulative backward sums.
A further modification to the procedure becomes apparent on noting
that, for example, the quantity S20 would provide the most efficient
measure that we could use to detect a step-change in the underlying
mean which occurred 20 time units ago. However, for our purposes we
would hope that any change of significant magnitude which occurred
would certainly be detected long before 20 time units after its
occurrence. In practice, then, we will only be interested in the first few
values of S. In a discussion of this method in the ICI monograph on
short-term forecasting (Coutie and others, 1966), the use of Sl9
S2, . . . , S6 is suggested. The control procedure is thus to specify
control limits =*= Lx, =±= L2— L6 for these sums and take the process
as being in equilibrium as long as all six St lie within their appropriate
253
limits. It is possible to calculate values for Lt,. . . , L6 so that each S

has some specified probability of falling outside the limits when the
process is under control. The six controls are actually used as a group so
that these individual probabilities are not very informative, particularly
since the values of Sx,. . . , S6 are not independent of each other. Thus
the problem of choosing theoretically a set of values for the L values is
extremely complicated. The procedure usually adopted is to express the
control limits Lt (i = 1, . . . , 6) as functions of i which involve some
simple parameters and then to choose the best parameters by a
simulation study. Obviously the values of L will increase with the
number of errors summed, i.e.
Ti L2 ^ L(y
A very simple form of control limit is obtained by taking Lt as a

linear function
Lt = Wo(i + h)
where a is the standard deviation of the errors and W and h are the
parameters to be chosen. To choose these parameters, simulation
methods can be used. Thus we generate two sets of data from a normal
distribution, or whatever distribution is appropriate, with the given o
and means zero and d, where d is a change in the mean of a magnitude
that should be detected fairly rapidly, e.g. d = 1.5 a. These data are
then used in the control scheme for a range of values of w and h. The
measure of control of clearest meaning here is the average run length
(ARL), which is the average number of observations after a change in
mean before the control system indicates a possible lack of control. If a
real set of data giving forecasts with average error close to zero is
available, these data and the data with d added to each observation can
be used in place of the simulated sets of data. Results from such a
simulation are given in the ICI monograph referred to above. By way of
example, suppose W = 0.4 and h = 3.4. Then when the situation is
under control the average run length is 9, so that with monthly data we
will unnecessarily review the situation once in 9 months. When,
however, a change in mean of magnitude 1.5 a occurs, a lack of control
will be indicated after about 2 observations.
The above describes the principle of the method, but for practical
operation a method can be devised by which only four quantities need
be recorded to use in the scheme. This suprising point is made in the
paper by Harrison and Davies (1964). The essence of the scheme is that
in looking for points which might cross the control limits one need only
look at the sum which is closest to the limits. A recurrence relation is
devised which enables this closest point to be identified at each stage,
without having to investigate points which Eire nowhere near the limits.
254
16.4 A Tracking Signal for autocorrelation
The tracking signals of the previous section were designed to indicate

the occurrence of any significantly large local bias. It is important,
particularly with stochastic models, to check the forecast errors for the
occurrence of autocorrelation between successive errors. Ideally the
errors will be independent. There is therefore need for a tracking signal
to indicate any large local autocorrelation as well as bias. The
autocorrelation of lag 1 is the natural one to study, but others might be
checked as well. A simple tracking signal for autocorrelation can be
based on the Von-Neuman ratio. This is normally used as a test for lack
of independence. It is based on the mean square successive difference,
i.e. on
n
2
5 = 2 (e,+ 1 ~ e,
n — 1 1=1
If the successive errors are independent, this has an expected value of 2a2.
The error variance a2 can be estimated by
1 n
S2 = - Z (et — e)2
n /= I
Thus the ratio K = 52 /S2 is approximately 2 if the e values are

independent. If there is a large positive autocorrelation, the ei+1 values
will tend to be close to the et values so 52 will be smaller. This can be
turned into a tracking signal by replacing all the above averages by the
corresponding discounted quantities. Thus
t-2
2 _ I
5 ^ (it — r ^t— 1 — r)
r- 0
where e t_ r is the exponentially weighted average of the e values at time

t — r. Thus, for large f, sufficient to let
t- I t- 2
2 b r
2 br
o o
we can define the tracking signal for autocorrelation by
t-2
2 br(et_ t- i
K - ill
K
t ~ t-i
Z) b {&t — r &t— r )
s2
r-0
255
As usual, numerator and denominator can be evaluated using recurrence

relations. Control limits for Kt are best found using simulations of both
independent errors and of errors showing autocorrelations of the
various magnitudes that one would wish to detect in practice.
References
Brown, R. G. (1959). Statistical Forecasting for Inventory Control. McGraw-Hill,
New York.
Brown, R. G. (1963). Smoothing, Forecasting and Prediction of Discrete Time
Series. Prentice-Hall. Englewood Cliffs, New Jersey.
Coutie, G. A., Davis, 0. L., Hassall C. H., Miller, D. W. G. P. and Morrell, A. J. H.
(1966). Short Term Forecasting. ICI Monograph No. 2. Oliver and Boyd.
Edinburgh.
Harrison, P. J. and Davis, 0. L. (1964). The use of cusum techniques for the control
of routine forecasts of product demand. Op. Res. Quarterly, 12, 325—333.
Trigg, D. W. (1964). Monitoring a forecasting system. Op. Res. Quarterly, 15,
271-274.
Chapter 17
Two-stage forecasting
17.1 Introduction
There are three basic aspects of applied forecasting: first, that of
analysis of past data and the construction of models for this; second,
the analysis of the use to which the forecast is to be put and again the
appropriate model building; and third, the bringing together of the two
to obtain the best forecasting method for the given application. Having
said this we must acknowledge that this is the ideal. At the opposite
extreme to the ideal, data are poured into one end of a computer
manufacturer’s forecasting program and whatever comes out at the other
end is used as the forecast. There are several factors that tend to
encourage people to take an approach that is less than ideal.
(a) Manufacturer’s programmes are frequently tied to specific fore-

casting methods and give the user very little opportunity to analyse
his data.
(b) The amount of work involved in the ideal approach is often
prohibitive. For example, it often occurs that a forecast of some
quantity is needed for several different purposes in a firm. The
ideal would be to use separate forecasts for each purpose since the
criteria will be different for different purposes, but the effort
required discourages this and a single forecast is probably used for
all purposes.
(c) There are only a few well-tried and understood methods of
forecasting, so there is a natural tendency to keep to these rather
than to launch out with tailor-made methods for particular types of
data and applications.
An implication of our failure to take the theoretically ideal approach

is that our forecasts are not the best that we could obtain. The question
thus arises as to whether we can improve our forecasts, without all the
work involved in the ideal approach to forecasting. In this chapter we
256
257
(a ) Ideal forecasting
Data Computer programme Forecasts
(b) Practical forecasting
Figure 17.1 The theory and practice of forecasting
will introduce the idea of two-stage forecasting. This seeks to improve

forecasts by means of a compromise between the more usual approach
and the ideal. The forecast obtained by some standard method is
regarded as the first stage in the forecasting process. We will assume
that the properties of forecasts obtained from this method, under stable
conditions, are known to us. The second stage consists of an adjustment
of this stage I forecast to obtain a new, stage II, forecast which is more
appropriate to the application or data on hand.
The basis from which the second-stage forecast can be constructed
include:
(a) the use of statistical properties of past forecast errors;

(b) the use of external information;
(c) the use of other forecasts;
(d) the use of different forecasting criteria.
By the use of any of these approaches we may seek to produce a

final forecast that is more appropriate to the situation but yet still uses
the standard methods to obtain the basic forecast. In fact we may
Data
'
Stage I Forecast
Past statistical information
External information
Stage H | Adjustment 1«
r Other forecasts
Specific criteria
| ForecasT
Figure 17.2 Two-stage forecasting

produce several forecasts for different purposes from the same stage I
forecast.
17.2 The use of forecast errors

A natural way to seek to improve forecasts is to see whether the
forecast errors still contain any structure that can be used. Let us
denote the stage I forecast of xt+h by xt + hi the error by et+h and the
stage II forecast by yt+h • By definition,
%t+h ~ ^t+h + ^t+h

If now by using the structure in the forecast errors we may obtain a
forecast, eit + h, of et+hi we can define a second-stage forecast by
yt+h ~%t+h +^t+h
The forecast error, zt+h, of the second stage is thus
A
^t+h ~%t+h yt+h

A
~ ^t+h ^t+h
by substituting from the above equation, It is clear that we will obtain
reduced forecast errors provided the forecast et+h is a reasonable
forecast of et+h. There are usually two features that we would like to
improve by the use of et+h. Firstly we would like to reduce any bias
that there may be in the stage I forecasts. We can do this completely if
the forecast et+ h is such that
E(et+h) = E(et+h)
for then E(zt+h) = 0. Secondly, we will want to reduce the forecast
mean square error, or variance if the forecasts are unbiased. Consider
the latter case. Denote the variances of zt+h, et+h and et+h by o2,a2
and k2 o2, respectively, and the correlation between et+h and el + h by
p. Then
ol = a2 + k2 a2 —2o ko p
For to be less than o2 we must have

p>k/2
Thus, for example, if the variance of et+h is the same as that of cf+h,
k = 1, then there is no point in introducing the second stage unless the
correlation between the forecast et+h and the actual error et+h is
greater than 0.5.
Let us now consider three different situations in which we can make
adjustments on the basis of evidence from previous forecasts.
259
17.2.1 Adjustment for bias

The methods of this section are appropriate where an examination of
the statistical properties of the forecast errors show some systematic
bias that can be used to adjust the forecast. Having observed such a
feature, a first step would be to examine the stage I forecasting method
to see if the adjustment might be made there without the use of
stage II. However, it is sometimes best to let well alone and correct the
forecast with a second stage. Such a situation might occur where the
stage I forecast is based on judgement. For example, forecasts of sales
are sometimes based on the judgement of a forecasting committee. For
reasons of strategy, possibly unconscious, the committee may over-
estimate sales. Rather than try to interfere with the functioning of this
committee, a better forecast could be obtained by reducing their
forecast by an adjustment based on an examination of bias in past
forecasts. One would thus have
Adjusted forecast = stage I forecast + estimated bias
Figure 17.3 Adjustment for bias
In graphical form the method is shown in Figure 17.3.

As an illustration of this technique, we consider two practical
examples. These examples also illustrate elaborations of the basic idea
of two-stage forecasting that may be used in other applications of
two-stage forecasting.
Example 1
The method discussed here was developed to provide forecasts of the
number of telephone calls arriving at a switchboard during different
half-hourly periods during a working week. The total time required to
answer these calls was also forecast. The data showed very large
systematic variations during the day and also between the days of the
\
260
week. There were also odd days where, for unforeseeable reasons, the
load was very heavy. The usual way to forecast such a situation might
be to fit some trigonometric model to the data, perhaps with the use of
some discounting procedure. Alternatively, some modified form of the
classical methods for seasonal adjustment could have been used.
However, both these methods would have required considerable
computation. The system required had to be used by someone without
any mathematical ability, and it had to give an answer by hand with
something like a minute’s effort. It had also to be very sensitive to pick
out the odd day with extra high or low loads. In addition it was
necessary for the system to follow changing patterns in the distribution
of load within the week. Such changes occurred sometimes during
different seasons of the year. Within this context any attempt to find
the best models or forecasting formulae with smallest mean square error
would have inevitably led to methods that were too complicated to
work. Instead, attention was focused on modifying the simplest
methods available.
The method devised was based on the technique of exponential
smoothing in its simplest form. As this technique was not appropriate
to data with periodic form, it was applied in two stages. An advantage
of the method was that the preliminary stage I forecast could be made a
week ahead and the final adjustment made on the actual day an hour in
advance of the half-hour being forecast. The method used in stage I was
to forecast separately for each of the 85 half-hourly periods in the
working week, using for each period only data from that period in
previous weeks. Thus exponential smoothing was applied to 85 separate
series. The nomogram given in Chapter 4 was used to simplify the
practice.
This method gave adequate forecasts most of the time. There were,
however, occasional days when the load was particularly high. This
produced a systematic positive bias in the forecast errors on such a day.
This bias was forecast one hour in advance by exponentially smoothing
the forecast errors from stage I as they arose during the day. Thus if
x10 was the stage I forecast, made the previous week, of the tenth
half-hour in the day and es was the exponentially smoothed error
obtained at the eighth half-hour, then at that time a revised forecast
would be
y i o = *io + e8
Such an adjustment was very effective on those days when something

special happened. It was found also that because of trends during the
year there was usually a small bias present on most days. If the stage I
forecast was always adjusted by a multiple, in this case 0.5, of the
smoothed error, there was found to be an overall 5 to 10 per cent
improvement in the forecast root mean square error.
261
Example 2
A large firm produces at the end of each year a planned sales for each
week of the following year. This plan was based on fitting a linear
regression model to log sales data and then making various adjustments
for holidays, etc. A study of several years’ data indicated that this was a
satisfactory model. The use of log sales in the model suggested that
errors be defined as percentage deviations from plan. Thus if pt is the
planned sales for week £, xt the actual sales and dt the percentage
deviation, then by definition
xt =pt{ 1 + df/100)
A study of past values of dt indicated that in any year there were runs
of deviations above and below zero and that there was an overall bias
away from zero. If we manage to forecast the future value, dt+h, by
dtfh, then it is reasonable that a revised plan will be given by
&t,h =Pt+h (1 +<2f,*/100)

Pt+h being known. As the plan is used to decide on capital investment
and manpower establishments at peak periods, it is clearly desirable to
seek to produce such revisions. A practical problem here is that the lead
time h, over which these decisions have to be made, may be so large
that our forecast dtfh is poor and the revised plan would have worse
properties than the original. This would require careful investigation in
terms of costs of errors in the plan. Assuming that dt h is sufficiently
good forecast, for our planning the deviation, d?+ h, of xt+h from
the revised plan is given by
d%h = (dt+h -dt,h)l(l+Jt'hllOO)
Table 17.1 gives some properties of the quantities d, d and d* from the
situation described. As the values d are obtained from deviations from
the yearly plans, there are no systematic changes with h. The forecasts
ct naturally get worse as h increases, however they have reduced the bias
in d to such a large extent that the mean square of d* is still
appreciably reduced over a nine-week lead time.
Table 17.1 Properties of ‘deviation from plan’ based on 19 weeks’ data
Lead time in weeks

1 5 9
Standard Mean Standard Mean Standard Mean
Mean deviation square Mean deviation square Mean deviation square
d 7.18 10.5 161.3 6.81 10.4 154.5 6.85 10.5 156.3

d 0.07 10.7 114.5 -0.39 11.2 123.6 -0.50 11.3 128.0
d* 0.19 10.0 100.3 -0.19 10.5 108.9 -0.28 10.6 112.0
\
262
It is not appropriate to discuss the forecasting methods used in this

particular example, except to point out that in this type of situation we
are often particularly interested in runs of deviations either above or
below zero and thus it may be advisable only to introduce a revised,
second-stage plan if there is some indication that such a run is taking
place. Or, alternatively, we may have a constantly used second-stage
revision of plan to allow for overall bias and, when appropriate,
introduce a third stage to allow for runs.
17.2.2 The use of autocorrelation in errors
In section 15.2 (viii) and Figure 15.9 we looked at plots to show

relationships between successive errors. If such relationships are found,
they can be used to forecast future errors and thus provide a
second-stage forecast. The most common relationship is an auto-
correlation between successive errors e. The simplest model for this is
the autoregressive model of section 7.1, i.e.
The parameter 0 is the autocorrelation of lag 1. This can be estimated
Table 17.2 Two-stage forecasting — autoregression adjustment

Autocorrelation of lag 1 = 0.48 = 0
Second-stage forecast yt = 5ct + 0ef_ i
1 2 3 4 5 6
(l)-(2) (2) + (4) <1> — (5)
xt xt e = (pet_ ! y
336.3 352.0 -15.7 -3.2 348.8 -12.5

339.5 363.6 -24.1 -7.5 355.8 -16.3
315.3 377.8 -62.5 -11.6 366.2 -50.9
359.0 384.7 -25.7 -30.0 354.7 4.3
364.3 382.3 -18.0 12.3 370.0 -5.7
353.0 347.6 5.4 -8.6 339.0 14.0
366.0 365.0 1.0 2.6 367.6 -1.6
369.0 372.4 -3.4 0.5 372.9 -3.9
363.0 367.7 -4.7 -1.6 366.1 -3.1
350.8 368.4 -7.6 -2.3 366.1 -15.3
Comparison
Mean 17.82 12.76
absolute
error
Example here uses x obtained by a regression method.

263
from past information or we can continually update our estimate as

illustrated in Table 7.5. Thus, given an estimate 0, we have for the
stage II forecast (see Table 17.2)
= +
Jt + 1 %t+ 1
17.2.3 Regression adjustment methods

Suppose that when we plotted a set of forecasts and observations on
the forecasts—observation diagram we obtained not the line of perfect
forecasts but some other perfect curve x = f(x), as in Figure 17.4(a). If
we now define a second-stage forecast by
y,+ h =f(xt+h)
then yt+h will equal xt+h and we will obtain perfect forecasts.
Obviously, in practice, we will not have anything remotely like a
perfect curve. It does, however, sometimes occur that the forecast-
observation diagram reveals a pattern of points through which we can
fit a suitable line by standard statistical methods. The most common
( a ) A perfect relation on the forecast-observation diagram
A A
( b.) A case of overestimation
Figure 17.4 Adjustment using regression

>
264
pattern corresponds to a bias whose magnitude increases with the

magnitude of the mean. Thus we have a regression line through the
origin but not at 45°, as is illustrated in Figure 17.4(b). The slope of
this line is estimated, using least squares, by
j3 = 2xx/2x2
Thus the second-stage forecast is
A
A /O A
y = P x
A study of adjustments, based on linear regression, applied to economic

forecasting is given in Theil (1958) and (1965).
17.3 The use of external forecasting information
We consider in this section the problem of adjusting forecasts to allow

for external information that can be expressed as simple mathematical
or statistical restrictions on the forecast. These methods are, for the
most part, based on the ideas in a paper by Zellner (1963), though the
approach is somewhat different. Consider a few examples.
Example 1
The stage I forecast of the output of a quarry may be greater than

the maximum amount that the mechanical handling equipment can
cope with, xmax say. In general we could write (see Figure 17.5a)
(X it x < xmav
Stage II forecast = {
I'^max if X > Xmax
A similar situation may arise where a lower limit xmin is set so that
if £ > x min
Stage II forecast =
if x < x min
or, alternatively, limits might be put at both upper and lower ends of
the expected range of the forecast value. These are shown graphically in
Figure 17.5(a) and (b).
Example 2
When forecasting sales with a growth model some scepticism about

the rapid growth forecast two years hence might be felt. If ‘expert
opinion’ thought xmax was a value that was unlikely to be exceeded,
then a compromise between x and xmax might be made when x
exceeded xm ax . In mathematical terms we might put
265
A A
(a ) Upper cut-off limits in stage U (b) Cut-off limits in stage II
(c ) Compromise limits in stage II
Figure 17.5 Adjustment using external information
if x < x max
Stage II forecast =
+
(1 6 )x max if x > x max
The parameter 0(0 < 6 < 1) would be made small if strong reliance was
put on the expert opinion rather than the forecasting method. If the
converse, 6 would be put close to 1.
Example 3
In some cases the external information can be put in the form of a

distribution for the future observation xt. That is to say that, apart
from knowledge of the stage I forecast, we have a detailed belief in the
possible values of xt in the form of a distribution function p(xt). The
expectation of this distribution, which we might call the a priori
distribution of xt, provides a single forecast of xt. Here, however, we
wish to use knowledge of the distribution to modify the stage I
forecast. For any distribution we may find a pair of values (L. U) such
that there is a small probability a of observations lying outside the
interval (L. U). We may choose L and U so that the probabilities are a/2
of observations lying below L and a/2 of observations lying above U.
For example, if p(x) is a normal distribution with mean p and standard
deviation a, there are probabilities of 0.025 of observations lying below
L = p — 1.96 o and 0.025 of observations lying above U = p + 1.96 a.
266
If the stage I forecasts lie in the outer tails of our a priori

distribution, as indicated by L and U, it seems reasonable that the
stage II forecasts should seek to pull the forecasts in towards the mean.
This could be done in many ways. The forecast could be set equal to L
or U if it exceeded the bounds, i.e.
L if x < L
Stage II forecast x if L < x U
or it may simply be a compromise between x and the bound L or U.
Thus
eLx + (1 —6l)L if x<L
Stage II forecast = x if L<x u
where 0L and 6 v are chosen constants. For eL = 0 this reduces to

our first rule. The forecasts are, of course, of the same form as those in
Examples 1 and 2, but here our choice of the limits is based specifically
on statistical considerations.
If the reasoning on which p(x) is based has more evidence on, say,
the upper range of values, we might choose U so that the probability of
exceeding U was very small and set 6 v also at a small value. This would
have the effect of making it very unlikely that the stage II forecast will
exceed U, which would reflect our confidence in this tail of the a
priori distribution.
Studies of the effects of cut-offs indicate that, provided the
information leading to the choice of L, U(and 6) is reasonable, the
stage II forecasts are significantly better than the stage I forecasts. The
information has to be positively misleading before the two-stage
forecasting ceases to be of value. This wrould occur, for example, if U
were some way below the true expected value of the future
observation.
17.4 Combining forecasts

There are often several reasonable methods of forecasting available for a
particular set of data. Rather than choose the best, which is often only
marginally the best, it is often worthwhile to seek to combine the
methods. If the methods use different features of the underlying
situation, then such a combination may produce better forecasts than
either forecast separately. A combination may be best achieved by
devising a new model that combines, the relevant features underlying
the forecasting methods considered. However, it will usually be easier
simply to combine the forecasts themselves. Suppose that we have two
unbiased forecasts, xl and x2 • A simple linear combination of these is
267
the most natural way of combining them. Thus the new forecast is
y = kx1 + (1 — k)x2
The new forecast y will be unbiased and will have a minimum mean
square error for suitably chosen k. If we consider the forecast errors of
the three forecasts, it is clear that
ey = ke1 + (1 — k) e2
using an obvious notation. If the forecast errors of x1 and x2 are
independent of each other, then E(ey) is minimum when
E(e\)
k
E(ej ) + E{e\)
(See Bates and Granger, 1959). In practice we would replace this by
ez2
k =
where e2 is a local estimate of the expected error squared. Newbold and

Granger (1974) found the average from the previous twelve forecasts
gave satisfactory results over a range of data. The exponentially
smoothed error squared provides another alternative for e2.
A somewhat different approach (Bunn, 1975), leads ultimately to a
form of scoring system for the choice of k. Thus if in the last n
forecasts xj has been the closest to the observed values on r
occasions, then
k = —
a
\ + + n
The parameters ax and a2 are chosen on the basis of one’s prior

assessment of the merits of the two forecasts. Thus if they were
thought to be equally good we would put ax = 1, a2 = 1 so that initially
k = y2.
Table 17.3 gives an illustrative example of these two approaches.
Note that in both methods k is an adaptive parameter that changes in
time to give the most weight to the most successful of the two
forecasting methods being combined.
As an example of a slightly different variation on this theme, let us
consider a case where the reliability of forecasts varied with time. In
forecasting future telephone traffic, staff from several hundred Tele-
phone regions’ were asked to forecast the percentage growth in
telephone traffic for a number of years into the future. The central
authority made an independent forecast for the country arriving at a
figure of a constant growth of, say, p per cent, over the years that were
forecast. It was felt that over short lead times of a year or two the
268
Table 17.3 Section of table showing combination of forecasts:

combined forecast luses the first method described, combined
forecast 2 the second
Combined Combined
Observation Forecast 1 Forecast 2 Forecast 1 Forecast 2
97 115 120 116.9 '115.7

102 106 94 101.5 104.7
96 79 74 77.2 78.6
100 75 108 87.2 77.0
84 68 87 78.2 73.3
MSE 14.6 14.3 12.7 13.9

observations
locally made forecasts would be moderately good, but for longer lead
times the p per cent, would be the best figure to use. It was also
assumed a priori that growth rates for all regions should be positive.
Using these assumptions the central authority adjusted all the regional
forecasts. Denoting the regional forecasts of the percentage annual
growth for a lead time of h years by xh and the central authorities
combined forecast by yh , the formula for the adjustment was
. fP if xh < 0
h
\ ahxh + (1 — ah )p if xh > 0
where a is some constant in the range (0,1). It is seen that this formula
is similar to those above with k replaced by ah . The effect of this form
for k is that for large lead times yh tends towards the value p. This
adjustment procedure, with values of a based on trials with past
forecasts, produced very large reductions in the average of the mean
square errors over all the regions.
17.5 The use of cost and other criteria*

Most forecasting exercises are based on a minimum mean square error
criteria. We have already referred to the situation where a single
forecast based on this, or some other specific criteria, is used by
different people or departments for different purposes. It may well be
that each of these has a different criteria for what constitutes a good
forecast. It may be that the costs depend on something different from
the error squared or that the criteria may be framed in terms of
minimizing probabilities rather than costs. In any such case, considera-
tion should be given to the possibility of adjusting the initial forecast to
produce a stage II forecast with better properties in terms of the
269
specific criterion. The aim of this section is to give some illustrative

examples of how this can be done. In all these examples we assume that
we know the distribution of errors in the original forecast. Such
information is usually available if the forecasting system has been in use
for some time or if a reasonable amount of data is available for the
investigation.
Example 1
Let the stage II forecast be of the form
yt = xt +a
where a is some constant to be chosen to minimize the expected cost of

the error in yt. Assume that past records of xt and xt indicate that the
forecast errors at stage I are normally distributed with zero mean and
variance o2. The assumptions of normality and zero mean are brought
in here to simplify the presentation; they are not necessary assump-
tions. The forecast error e of the stage II forecast will normally be
distributed with mean -a and the same variance.
Stage I error
Stage II error
Figure 17.6 The effect of an addi-

tive adjustment
Many possible forms of cost function can be considered. It is clear

that if the cost function and the error distribution are symmetric, with
minimum cost when the error is zero, there will be no need to adjust
the original forecast. There will only be a saving if the cost function is
non-symmetrical. In such a case a biasing of the forecast, by putting
a =£ 0, will move the range of errors with highest probability into the
region of the cost function involving lowest costs. It can be shown that
for data based on a normal distribution and appropriate linear
forecasting formulae the optimum predictor is obtained by such a
procedure (see Granger, 1969).
Cost functions formed of linear sections are the simplest non-
symmetric functions to manage mathematically. For example, the cost
functions in Figure 17.7(a) and (b) are broken into four sections
270
(a ) Cost
Error
-r 0 s
(b) Cost
Figure 17.7 Some illustrative cost func-

tions
representing minor under-or overprediction and major under- and

overprediction. This type of cost function can be fitted to actual cost
data as a reasonable approximation without requiring a great deal of
information.
It is not proposed to go into mathematical detail in this section. Let
us simply examine the general idea of the method. Let us denote the
quantity (e + a)jo by u; this will have a standard normal distribution,
iV(0,l), with cumulative distribution denoted by (u). That is to say,
<f> (uQ ) gives the probability that u is less than or equal to uQ. If we have
the cost function shown in Figure 17.7(a), then we can calculate the
probabilities of the various costs occurring. Thus there is a cost of
amount a if u < a — r/a which has probability $ (ce — r/a). All the costs
and probabilities are given in Table 17.4.
The total expected cost can thus be found by multiplying the costs
by their probabilities and summing. This cost is a function of a, so a
can be chosen to minimize the cost. Doing this leads to the equation
a +s
(a — b)(j> + (b — c)0 + (c — d)<j> =0
o
where 0(u) is the ordinate of the standard normal curve. It will be seen
that if the cost function is symmetrical with a = d, b = c and r = s the
solution of the equation is a = 0 as we would expect. For the
non-symmetrical case with b = c = 0 the equation becomes
= d/a
271
Table 17.4 Costs and probabilities corresponding

to Figure 17.7(a)
Cost Range of u Probability
a—r a—r
a u< <f>
o o
a—r a a—r
*1 “
-4>
o o o
a a+s a+s
-<u< 4> 4>l ^
o o o
a+s a+s
<u 1 — 4)
o o
The implication of this is that the greater d/a the greater a as is to be

expected.
The same procedure can be applied to cost functions such as those in
Figure 17.7(b) and (c) and also where the costs depend not just on the
magnitude of the error but on the values of x and y. These methods
tend to lead to rather involved equations, but they can provide a useful
method of adjusting forecasts to meet more realistic cost criteria.
Example 2
Let us now examine an example where the criteria of interest is not a

cost but a probability. Suppose we wish our stage II forecast to be such
that the probability of the future observation being more than 10 per
cent, out is as small as possible. That means to say that we want
P = Prob (xt < 0.9 yt or xt > 1.1 yt)
to be small. In general we would require to get yt to minimize an

expression of the form
P = Prob (xt < byt or xt > ayt)
From the distribution of the stage I error we can find the distribution
of xt conditional on xt. For example, xt might be distributed as N(x t, o2 ).
Thus we can find P as a function of both yt and xt. We may then choose
y to minimize P. If we denote the distribution of xt given xt by f(xt I xt)
this exercise gives
f(byt 1 xt) = a
f(ayt\xt) b
as the relation giving minimum P. As an example, if the stage I forecast
272
Table 17.5 Examples of stage II fore-

casts
a = 1.1, b = 0.9
xt 02 yt xt a
2
yt
10 10 16.20 100 1 100.10
20 10 24.15 100 4 101.58
50 10 51.93 100 5 110.24
100 10 100.99 100 50 208.38
500 10 500.20 100 100 366.76
a = 2.0, b = 0.5
xt O2 h
100 1 80.00
100 50 80.46
100 200 81.80
100 1000 88.37
100 2000 95.48
100 4000 107.50
errors are normally distributed with zero mean and variance o2,
left-hand side of the above equation becomes
exp — TlUWt ~ *r)2 - {ay,- xt)2}

2 (T
Hence we find
„ „ a a + b
x] + 2o2 log l (a + b)
ba-b
Some illustrative values of yt are given in Table 17.5. It will be noted
that if o21x2 is small, i.e. we have good forecasts in percentage mean
square error terms, then the above result approximates to
y = 2x/(a + b)
Thus for a + b = 2, as in the symmetrical example we started with,
y = x. If, however, we allow for a higher percentage of error above
y, then a increases and so y moves below x.
References
Bates, J. M., and Granger, C. W. J. (1959). The combination of forecasts. Op. Res.
Quarterly, 20,451—468.
273
Bunn, D. W. (1975). A Bayesian approach to the linear combination of forecasts.

Op. Res. Quarterly, 26, 325—330.
Granger, C. W. J. (1969). Prediction with a generalised cost of error function. Op.
Res. Quarterly, 20,199—207.
Newbold, P., and Granger, C. W. J. (1974). Experience with forecasting univariate
time series and the combination of forecasts. J. Roy. statist. Soc. A, 137,
131-146.
Theil, H. (1958). Economic forecasts and Policy. North Holland Publishing Co.,
Amsterdam.
Theil, H. (1965). Applied Economic forecasting. North Holland Publishing Co.,
Amsterdam.
Zeilner, A. (1963). Decision rules for economic forecasting. Econometrica, 31,
111-120.
>
Chapter 18
Problems in practice
18.1 Introduction
The aim of this chapter is to discuss some problems that arise when
forecasting is used in practice. In this discussion we will not attempt
anything like a full coverage of the field. Instead, we will examine a
number of selected examples. The examples have been chosen to
illustrate one main theme. This theme is that it is incorrect to regard
forecasting as an exercise that can be done on its own without reference
to the use to which we put the forecasts. All too frequently this is
exactly what is done. Having produced a forecast it is regarded as a
number that can be used for any situation where a forecast is required.
We might say that in such a case the forecast and application are
'decoupled’ from each other. Such decoupling often occurs where a
firm has a ‘forecasting program’ on its computer used simply to
generate forecasts which are then given to anyone who needs them.
There are several possible sources of trouble in adopting this type of
approach to forecasting. These are discussed and illustrated in this
chapter. In summary, incorrect decoupling of forecast and application
can lead to:
(a) the use of forecasts derived from inappropriate forecasting criteria;

(b) the use of inappropriate forecasting parameters;
(c) the occurrence of forecasts that may be very accurate as forecasts
but are nonetheless of little value when used in the application;
(d) the occurrence of self-defeating forecasts;
(e) the creation by the forecast of unfortunate side effects in the
situation;
(f) the occurrence of self-validating forecasts.
There are situations where decoupling is quite possible and satis-
factory. There are others where a ‘partial decoupling’ is possible. In
these latter situations an initial forecast is obtained without reference
274
275
to the application and is later modified to suit a particular application.

These topics were dealt with in Chapter 17 on ‘two-stage forecasting’.
The alternative to decoupling and partial decoupling is the integra-
tion of the forecasting aspect with other aspects of a particular
application to produce a ‘system’ for doing the job required.
In this chapter the aim is to illustrate some of the phenomena
referred to in (a) to (f) above.
18.2 Practical forecasting criteria — a stock control example

The following example is based on a paper by Trigg and Pitts (1963).
The reason for selecting this example is that it illustrates how
forecasting can be closely linked to an application and also how a
forecasting parameter can be chosen to optimize a criterion related
directly to the application.
The basic problem of stock control is that of giving the best service,
to those who wish to draw on stock, at the lowest investment in stock.
To give the best service it is necessary to make the probability of
running out of stock as small as seems reasonable. In this example a
probability of 0.025 (2V6 per cent) has been chosen. To achieve this the
total stock in hand and expected deliveries to the stock should be
greater than the forecast demand by a safety factor which is called the
‘safety stock’. If the demand from the customer each week has been
t • • ^ ^j ^ X ^ then the forecast demand for next week using exponential
smoothing is
oo
xf = (l — a) Z arxt_r (1)
r-0
One way of choosing safety stock is to let it be several weeks demand

plus some constant amount q. How many weeks are used will depend
on the time taken between ordering for stock and receiving these
orders. To simplify this example, we assume this time to be one week
and hence the safety stock is one week’s demand plus q, i.e. xt + q.
Suppose that the amount ordered from the producer at the end of week
t is Pt and the amount of stock at that time is St. The change in stock
over the last week is St — St_ j, which must equal the deliveries based
Figure 18.1 Model for the stock control

example
276
on the previous week’s order Pt_ j minus the amount drawn xt. Hence
St St_ i — Pf~i oct (2)

It is also necessary that our latest order for replenishment Pt must be
such that with our present stock St it must meet next week’s demand,
forecast as xt9 and provide for the safety stock. Hence
Pt +St=xt + (xt+q) (3)

Changing the suffix to t — 1 and subtracting gives
Pt-Pt-i =St_l -St+2(xt-xt_l) (4)

Substituting from equation (2) in (4) gives
Pt =xt + 2(Sct—xt_l) (5)
Substituting this in equation (3) again gives
St= — xt + 23ct_i +q (6)

Equation (6) thus relates the stock held to the latest demand, the
previous forecast and the size of the safety stock.
We now consider two possible situations. In the first we assume that
the demands are independent and follow a normal distribution with
variance a2. As xt is independent of xt_l, the variance of Sti a|, is
given by
a2 = a2 +4 Var(5cf _ x)
using equation (6), which is
0?=a2 1+4 (7>

( rrf)
using equation (1). To achieve our 2V& per cent probability of ‘stock
out’, we would make our safety stock equal to 2 ar We can then make
this small by making os small; this would be done by making a large
and, in fact, if we could assume a global constant mean model we
would use the ordinary mean as the forecast in place of oct. However, in
the usual stock control situation such an assumption is often
unrealistic. It is more reasonable to expect trends to occur in the data,
at least locally. The lag of the forecast xt behind an upward trend
xt = ftt would soon lead to the safety stock being used up. To avoid
this, we add to the safety stock an amount Dm which is the maximum
depression in stock produced by such a trend.
Table 18.1 shows some illustrative values of the various quantities in
this situation calculated by using equations (2) and (3). It is seen from
this table that the system is such that after the fourth week the orders
delivered overtake the trend in sales, so that the depletion in stock
277
Table 18.1 The effect of trend on stock
Week t 0 1 2 3 4 5 6
Stock at t — 1, St_! 10 10 9.00 8.40 8.12 8.10 8.28

Orders delivered Pt_! 0 0 1.40 2.72 3.98 5.18 6.34
Sales xt 0 1 2.00 3.00 4.00 5.00 6.00
Stock at t, St 10 9 8.40 8.12 8.10 8.28 8.62
Orders made Pt 0 1.4 2.72 3.98 5.18 6.34 7.48
Forecast xt 0 0.2 0.56 1.05 1.64 2.31 3.05
Depletion Dt 0 1.0 1.60 1.86 1.90 1.72 1.38
reaches a maximum at week 4. To find the depletion Dt in stock at

week t we define Dt by
Dt=S0-St (8)
which, by equation (6), taking the initial stock S0 as just q, gives
Dt = xt — 2xt_ i
Substituting the model for xt in xt gives
xt = (3t — af3( 1 — af) 1(1 — a)

Thus on substituting and sorting out we find
Dt = /5{ 2 - (1 - a)t - 2af} /(I - a) (9)
Figure 18.2 shows a rough sketch of this type of curve. It is seen that
there is a maximum depletion Dm in stock after a time tm (to the
nearest integer), but that after this the stock in fact starts to increase
and the system overcompensates for the trend. In theory this looks bad,
but in practice the type of trend we are considering here is simply a
local phenomenon. If it appears to have become a permanent part of the
situation, the forecasting formula would be changed to a trend-
correcting form. The intuitive explanation of the form of Dt is that
initially in the transient state, the forecast does not follow the trend, so
Figure 18.2 Depletion in stock —sketch

of shape of curve
278
the orders do not keep up with the demand. After a time the forecast
increases at the same rate as the trend but with a fixed lag behind it.
Though in this steady state the forecasts are lower than the sales, the
occurrence in the ordering rule of equation (5) of the term 2
— xt-i), which has expectation 2/3, means that the orders will on
average be for more than next week’s sales and thus, in the steady state,
there will be a continual increase in stocks. The calculation of Dm from
Dt is best done by evaluating D1? D2, etc., until the largest value is
found. If there were no random variation in sales and we could
reasonably guarantee that no local trends would have a slope greater
than |3, then a safety stock greater than Dm would ensure that no ‘stock
out’ is likely to occur.
If we now allow for both random variation and the possibility of a
trend, we see that it is reasonable to set the safety stock at 2 + Dm .
Referring to equations (7) and (9) it is clear that both these quantities
depend on the choice of a, if a is increased then os decreases and Dm
increases, and vice versa. Thus there is, in fact, a value of a for which
the safety stock is minimum. This value, for example, is approximately
0.95 for a/|3 = 10. We thus have here an example where a forecasting
parameter is chosen, not with reference to getting best forecasts in any
mean square error sense but with reference to the particular application
in which the forecast is used.
18.3 The forecast as part of a system — a stockholding example
In this section we will look a little more at the idea of stockholding. We

saw in section 18.2 that the choice of forecasting parameters can be
related to the quantities required of a stock control system. However, if
one did aim at optimum forecasts for the system one could treat the
forecasting and stock control as separate exercises. Very often in
business situations the very data one uses for forecasting is affected by
somebody else’s efforts at stock control. In section 18.2 we looked at a
stock control application from the point of view of the person
concerned with the operation of the stock control system. Let us now
look at that same situation from the point of view of the person or firm
who supplied the stock. For this person the quantity Pt is the order
that comes to him at time t and goes out later as his sales. For the sake
of concreteness le us regard him as the manufacturer and then the
stocks S are those held by the retailer and xt refers to the retailer’s
sales. If the public is only going to buy, say, 1,000 items from the
retailer it is clearly best for the manufacturer to only manufacture
1,000 items. However, to be in this happy state he would have to
produce a perfect forecast of the values. Unfortunately, he is usually
not able to do this and in fact as he does not have access to the
279
retailer’s data he cannot forecast x at all. The only information he has is

the data of his sales which is the sequence of the P values. One would
like to think that this gave him relevant information about future sales
to the public so that he can plan his manufacturing accordingly, Let us
examine, therefore, the relation between the sales to the public and the
manufacturer’s sales. Assuming the same simple situation envisaged in
section 18.2 we have equation (5) of that section as
Pt = xt +2(xt — xt_1)
Let us notice a number of different facets of this equation.
(a) Suppose we are using exponential smoothing with
xt = axt_ j + (1 — a)xt
we then have
Pt = (3 — 2a)xt — 2(1 — a)xt_ x
If for simplicity we assume that xt and xt_ x are independent, then

the variance of Pt is given by
Var(Pf) = (3 - 2a)2 Var(x,) + 4(1 - a)2 Var(xf_ x)

It is sufficient here to simply note that the coefficients of Var(xf)
is bigger than one. If a = 0.8 it is (1.4)2. Thus the manufacturer’s
sales have a greater variability than the retailer’s and it is
consequently more difficult for him to produce forecasts of future
requirements. The reason for this additional variation lies in the use
of a safety stock whose size depends on forecast sales.
(b) Let us now consider the case where sales show a steady rate of
growth or decline and that the retailer makes a simple forecast of
the form:
xt = cxt
Thus if he expects a 5 per cent growth rate he would use c = 1.05.

In this case we have
Pt = xt + 2 c(xt — xt_x)
Thus if x is in fact increasing Pt > xt9 irrespective of c, and if x is
decreasing Pt < xt. The consequence of this is that if sales are
increasing the manufacturer will overestimate, in his forecasts, the
total amount to be sold ultimately to the public If sales are
decreasing he will underestimate. The amount of overestimation or
underestimation will increase with the rate of change of x. An
example of this type of behaviour is given in the paper by Kuchn
and Day (1963), where the term ‘acceleration effect’ is used to
describe the phenomena.
280
(c) Suppose now we consider what happens if there is an oscillation in

the sales x. For simplicity let us suppose that the sales x follow the
sequence b + m, b — m, b + ra, b — m,. . . . With the forecasting
formula of (b) a little arithmetic will show that P would follow the
sequence b + (1 + 4 c)m, b — (1+4 c)ra, b + (1 + 4 c)ra, .... Thus
the height, or amplitude, of the oscillation has increased by a factor
of (1 + 4 e).Thus the man who always assumes that next month
will be the same as this, c = 1, will have magnified the oscillation by
a factor of five. There is a further feature here worth noticing, and
that is that as Pt is actually sold at time t + 1 the manufacturer will
be delivering b — (1 + 4 c)m to the retailer when the retailer is
delivering b + m to the ultimate customer. This type of occurrence
can cause the retailer to worry and play safe by putting up the
value of c next time he orders. Thus in our example there is a
change in amplitude and also a lag between the peaks of x and
those of P.
In (a) to (c) above we have considered a very special example with a

lead time of one and some particularly simple safety stock and
forecasting rules. These examples, however, illustrate some very widely
occurring phenomena. Figure 18.3 summarizes the phenomena dis-
cussed. Often there is more than one middle-man in the system. For
example, in areas of clothing manufacture the clothes are bought from
a shop, which carries a stock. The shop buys from a wholesaler, who
carries a stock. The wholesaler buys from the manufacturer, who carries
Randomness
(a)
amplified
(b) Trends
exagerated
Oscillation
(c) /W\
AM amplified
Change
_y\
delayed
Figure 18.3 Some effects of stockholding

281
a stock of clothes and also a stock of material. This manufacturer may

buy from a materials manufacturer who may in turn buy from a
manufacturer of the appropriate fibres. Each, of course, will have his
own stocks. Any change in the sales of clothes in the shop will slowly
work its way back through the system, being magnified each time. In
one classic example of this the sales to customers changed by something
like 5 per cent. It took 6 months for this change to work its way
through the chain and the change was amplified to 50 per cent by the
time it reached the producer of raw materials.
A more involved study of the single-stock situation is given in the
paper by Leicester (1963), which looks at the sales to and by retailers
in the shoe trade. This has a number of complicating features. A
retailer’s subjective expectations of the state of the market influence his
orders from the manufacturer. In this situation the delivery time is a
varying quantity also depending on the state of the market. The
amount of safety stock is dependent on this delivery time or rather on
the retailer’s estimate of what it will be. A consequence of the
amplifications and variable lags produced by this system is that when
the sales of shoes by retailers show a steady upward movement,
manufacturers’ sales show a strong cyclical oscillation of booms and
then recessions.
Having been warned of the effects that stocks may have on orders
from stockholders, what should the forecaster do with such order data?
There are a number of actions open to him. He may seek to obtain data
on sales to the ultimate customer which will provide the forecasts
necessary for manufacturing. If this is not possible, he must take his
forecast orders as information upon which the decision as to how much
to manufacture will be based. To reach this decision he will need to
develop a realistic model of the total system that generated his data. If
the system is fairly simple, he may even be able to use his order data to
estimate the sales-to-customer data, e.g. in (b) above knowledge of the
sequence of the P values provides information about the sequence of
the x values.
18.4 Forecasts and crystal balls — a scheduling example

A very common error in the use of forecasts is to treat the forecast in
the same way as one would treat the true future observation, if only
one had a crystal ball in which to see it. However, the number that
comes out of the crystal ball is a known constant, whereas the number
that one obtains from one’s efforts in statistical forecasting is
essentially the value taken by a random variable. This distinction
between a constant and a random variable must be clearly realized if
forecasts are to be correctly used in practice. The object of the
following section is to emphasize this point by considering examples in
282
which the replacing of a constant value by a forecast can completely

change the situation.
Let us begin by considering an idealized example to show that a
forecast treated as a ‘crystal ball’ forecast can lead to a different result
to the one which would be obtained if it is correctly treated as a
random variable. The problem is to decide whether to increase plant
capacity by one or two units (C = 1 or C = 2), assuming that the future
additional demand, D, will be for either one or two units.
Model
Possible demand D 1 unit 2 unit
Average weekly profit

if C= 1 10 10
if C = 2 5* 20
(*penalty for surplus plant)
(a) Suppose the crystal ball forecast demand is D = 1 unit. If C = 1

then the profit is 10, if C = 2 the profit is 5. Hence the decision is
C = 1 for maximum profit.
(b) Suppose that the forecast, D = 1, is based on estimated prob-
abilities:
Prob(D = 1) = 0.6
Prob(D = 2) = 0.4
The expected profit
if C = 1 is 10 x 0.6 + 10 x 0.4 = 10
if C = 2 is 5 x 0.6 + 20 x 0.4 = 11
(multiplying the profits by the probability of showing these

profits). Hence the decision is C = 2 for maximum profit. Thus
knowledge about the random aspects of the situation leads to a
different decision being obtained from the forecast I) = 1.
Let us now consider a less idealized example in more detail. A very

common practical problem is the choice of the number of ‘servers’
required to give a good service to ‘customers’. Common examples of
this are the number of cash desks in a supermarket and the number of
telephone operators manning a switchboard. We consider here the
specific problem of scheduling how many operators are required to man
a switchboard. To answer this type of question, we require first of all
information about the criteria of good service and, secondly, a model of
how the particular queueing system functions. As a measure of service
we take the probability that a caller can get straight through to an
283
operator. For a numerical example we will say that service is

satisfactory if this probability is greater than 0.99 or, conversely, that
the probability of the caller being delayed is less than 0.01. The factors
effecting the service the caller obtains are:
(a) the number of operators N;
(b) the statistical properties of the length of time taken by an operator
to deal with a call;
(c) the statistical properties of the way in which calls arrive;
(d) the way in which calls queue up when all operators are engaged;
(e) the total load on the switchboard, 6.
In this example 6 is measured in units of 100 seconds of contact
between operators and callers. Over many years of study the actual
behaviour of telephone systems has been described very successfully by
a number of fairly simple statistical models. The model used for the
numerical calculations in this section is the Erlang C model which is
one of the simplest of these models. The principles of the discussion,
however, apply to all models. Morse (1957) explains these models in
detail.
Table 18.2 Maximum loads for Prob (delay) < 0.01. Units are 100 seconds of
operator contact per half-hour for the Erlang iC model
Number of operators M 1 2 3 4 5 6 7 8
Maximum 6, (0M) 0.1 2.6 7.7 14.4 22.5 31.5 41.2 51.7
Number of operators M 9 10 15 20 25 30 40 50
Maximum 6, {6M) 62 73 133 197 265 335 479 627
If we use the Erlang C model the probabilities of callers being

delayed can be calculated for any given number of operators and any
specified load. Table 18.2 gives the type of result obtained from such
calculations. It is seen, for example, that, with four operators and the
criteria of good service given above, the highest load that could be dealt
with is 14.4 units. When the tables are used in practice, they are in fact
used the other way round. For example, the question might be ‘How
many operators are required to give good service when the load is 30
units?’. From the table the answer is seen to be six operators.
In practice the load on telephone switchboards varies considerably
during the day and also between the days of the week. Figure 18.4
shows the type of variation that occurred in one case. It is thus
necessary to have different numbers of operators available at different
times during the day and during the week. If we knew, for example,
what the mean load was going to be for each half hour during the week,
284
Units
Figure 18.4 Variation in load on switchboard
we could use Table 18.2 to schedule the number of operators to man

the switchboard throughout the week, so that callers are assured their
99 chances out of 100 of getting directly to an operator. Thus if we had
our crystal ball to see what next week’s load would be, there would
be no real problem. However, what we in fact possess are forecasts
based on observations on previous week’s loads. What is usually done in
this situation is to take the forecast, say 25 units for Monday
9.00—9.30am, and use it in the tables to decide how many operators
are required, in this case six, exactly as though it were the true value.
But our forecast will probably be wrong and the true value may turn
out to be 35 units for which we should have had seven operators. Thus
we have one less operator than is needed to give good service, and this is
a direct result of our forecasting error.
An immediate question that arises is whether we can use the
statistical properties of the forecasting error to obtain the probabilities
of being in error in the numbers of operators. This can in fact be done
either by a detailed study of the queueing model or by an approximate
method. We will indicate briefly the approach of the approximate
method as it, or some modification of it, can be used to study the
effects of forecast errors in any situation where the foreccists are used in
conjunction with tables such as Table 18.2.
The idea of the calculation is indicated by Figure 18.5. It is assumed
that the forecast 6 is distributed about the true 6 with a normal
distribution of variance o2. It is also assumed that over small ranges of
3 or 4 times o the table can be approximately represented by a set of
equal intervals of width C on the axis of 0, each interval corresponding
to a different number of operators. It is clear from Figure 18.5 that if
we knew the position of 0 we could find immediately the probabilities
285
Figure 18.5 Effects of forecast errors on reading a table
Table 18.3 Approximate probabilities

of being K servers in error, Prob(X =
—k) = Prob(A = k), for the Erlang C
model
r = C/a
K 0.5 0.75 1.0
0 0.195 0.286 0.369

1 0.173 0.221 0.241
2 0.121 0.102 0.067
3 0.066 0.028 0.008
4 0.029 0.005 0.000
5 0.010 0.001 0.000
6 0.003 0.000 0.000
7 0.001 0.000 0.000
8 0.000 0.000 0.000
of being any specified number of servers in error. For example, the

probability of being understaffed by one would be p_ x, the area of the
appropriate region of the distribution curve. However, the position of 6
in the interval C is not known. It seems reasonable that it is equally
likely to be anywhere in the interval. If it is reasonable to make this
assumption we can obtain, using the idea of conditional and marginal
distributions, the probabilities of being, say, understaffed by one,
without assuming 6 known. Table 18.3 gives the results of this type of
calculation. The probabilities of having the wrong numbers of servers is
seen to depend on the ratio r = C/o. This is a measure of how far the
distribution of the forecast is spread over the intervals of the table.
Thus with r = 1 about 99 per cent of the distribution covers four lines
286
of the table. In the type of application in which this problem arose this
corresponds to quite a good forecast, but yet it gives a probability of
0.24 of being understaffed by one. With larger forecast variability
considerably worse errors are seen to occur.
As the probabilities of being understaffed and over-staffed are
equally likely, one might think that no great damage has been done.
Unfortunately, overstaffing improves the system very little whereas
understaffing has disastrous effects on the quality of service given. For
example, for the case where we scheduled six operators and had a load
of 35 units, which required seven operators, the delay probability is
0.03. If the load had been 50 units, this would go up to 0.07. Thus
though the probabilities of error in the numbers of operators is
balanced between understaffing and overstaffing, the effect on the
probability of delay is to increase it above the value of 0.01 used in the
original table.
The fact that the load is forecast rather than known completely
changes the meaning of the tables. Table 18.4 gives approximate values
for the actual delay probabilities when forecast values of load are used
in Table 18.2. The basis for the derivation of Table 18.4 was the use of
the probabilities of Table 18.3 together with information from
queueing theory on the effects of under- and overstaffing. It is clear
from Table 18.4 that if one has a small load for which two of three
operators are adequate, then large forecasting errors can lead to very
poor service. One way to improve service is obviously to improve the
forecasts. Whether or not this can be done, the obvious answer to the
difficulties that are produced by the use of forecasts in the tables is to
use the forecasts in a different way to the way one would use the true
load if it were known. One simple method of doing this would be to
use, in the tables, not the forecast but an adjusted forecast. The
adjusted forecast would be larger than the original forecast, thus cutting
down on the probability of understaffing and putting up the pro-
bability of overstaffing; this in turn will decrease the actual probability
of delays occurring.
We have discussed above a situation where a forecast, 0, is used to
choose between a number of alternatives, M = 1, M = 2, etc. The basic
objective is to use the forecast to make the best decision M in terms of
some criteria, here the delay. We have discussed this in some detail as it
is an example of a very common type of forecasting situation.
In summary, the lesson of this example is that we cannot treat
forecasts in the same way as we would treat the true values. The effect
of the random variation in forecasts can, however, be studied and
appropriate adjustments made once its effects are understood.
Before leaving this example, it is perhaps worth noting the possibility
of extending to other applications the use of the type of information
shown in Figure 18.5. Suppose, for example, that a study of a forecast
287
Table 18.4 Approximate actual probabilities
of delay when using tables
(a) Probability of delay given in table = 0.01
Number of r = Cjo
operators 0.5 0.75 1.0
2 0.116 0.074 0.046

3 0.077 0.041 0.024
4 0.059 0.027 0.015
5 0.044 0.020 0.012
6 0.036 0.016 0.010
7 0.027 0.012 0.008
8 0.024 0.011 0.008
9 0.020 0.010 0.007
10 0.017 0.009 0.007
(b) Probability of delay given in table = 0.1
Number of r = C/a
operators 0.5 0.75 1.0
2 0.185 0.149 0.123

3 0.172 0.132 0.106
4 0.151 0.113 0.092
5 0.138 0.102 0.083
6 0.131 0.096 0.079
7 0.123 0.091 0.076
8 0.113 0.084 0.071
9 0.108 0.081 0.070
10 0.103 0.078 0.068
errors leads to an estimate of the distribution of future demand for

some product given the forecast, e.g. that shown in Figure 18.6.
Suppose that on the basis of this certain production capacity, P, is
invested in. If the demand lies in the region A there will be idle capacity
under normal working, if in B overtime will be required, if in C
additional production facilities will be required and if in D the demand
will not be met. If the costs of these alternatives can be analysed then,
the optimum production capacity P can be chosen. A paper by
Thomeycroft, Greener and Patrick (1968) details such an analysis.
18.5 The forecast as part of a system — self-validating and self-

defeating forecasts
There are a number of situations where forecasts turn out to be very

accurate or completely wrong, not because of the skill or otherwise of
288
Figure 18.6 Consequences of deviation from forecast
the forecaster but because of the nature of the situation being dealt
with. Let us briefly consider some examples of these effects.
(a) When a new product is launched, sales figures may show a nice
steady increase. One is tempted then to forecast a continuing
increase. There have been cases where this has been done only to
find a very sharp levelling off at a later stage. The explanation has
been that there has been a potential market for say 1,000 sales per
week, but when the item was first introduced the production
capacity was only geared to 500 a week. The production capacity
has then been increased steadily. Thus until the 1,000 items/week
figure was reached the sales equalled the production and forecasts
based on the trend were naturally correct.
(b) The use of forecasts in setting salesmen’s targets produces some
interesting phenomena. Suppose the forecasts were made a target
and no incentive was provided for passing the target, though one
was provided for reaching the target. In this case one would expect
the salesman’s sales to equal the target and hence the forecast.
Alternatively, if there was a fixed pay up to the target and good
bonuses for exceeding the target, then there is a fair chance that
the forecast will be exceeded. In essence one really needs to
distinguish in this situation between a forecast of the natural
market, the target given the salesmen and the forecast of the future
sales, allowing for the influence of the target and the pay structure.
(c) In (b) we have a situation where the value forecast actually
influences the future sales. Another example would be where the
sales staff takes one look at the disasterous forecast of sales and
takes drastic action to change the situation. In this case any
resulting forecast error is not a measure of bad forecasting but
partly of the success or otherwise of the actions taken.
(d) An interesting example is given by Forrester (1961), which, though
there based on a simulation experiment, certainly happens in
practice. This is the situation where an initial use of a seasonal
289
forecasting method either creates a seasonal that need not exist or

exaggerates one that does exist. Thus at the peak period forecast,
advertising is at its maximum and the salesmen are working all
hours. The result is that sales are naturally at a peak and will
probably be higher than forecast. Next year’s seasonal forecast thus
has a tendency to show a greater seasonal effect than this year’s. As
a seasonal forecasting method will forecast seasonal elements even
if none exist in the data, it is possible by the above policy to induce
a seasonal sales where none need exist. Once again we have an
undesirable feature produced by using the forecast in a situation
without due regard to the decision nature or complexity of the
situation.
(e) Another example is related to oscillation features in situations
where production is geared directly to forecast sales. It can happen
that the sort of magnification that was seen in section 18.3(b) and
(c) allows the usual random variation in the forecast sales to cause
larger swings in production. A consequence of this can be that at
some periods production cannot meet demands. Thus sales are
artificially lower than they should be or than they were forecast to
be. We are thus misled in future forecasts by the type of
phenomena referred to in (a).
Once again the message of these examples is that forecasting cannot
be treated in isolation. It must be considered in relation to the
practicalities of the situation for which the forecast is needed.
18.6 Practical forecasting — a postscript

In the previous sections we have looked at particular problems arising in
the practice of forecasting. Let us finish this chapter, and this book, by
taking a brief look at some of the main issues that face the forecaster in
developing a practical forecasting system. As we are concerned with
statistical forecasting, we can leave aside aspects of administration and
implementation and concentrate on those questions that are relevant to
the use of statistical models and methods for obtaining and using
forecasts in practice. Each question will be discussed briefly and
reference is made to the sections of this book that are most relevant.
(a) Why is the forecast required ?

This is not a statistical question, but a knowledge of the answer is
essential in providing a framework in which to consider the more
statistical aspects of forecasting. Questions raised on the appropriate
criteria to use (e.g. Chapter 3 and sections 17.5 and 18.2) or on the
level of complexity of the methods appropriate (e.g. Chapters 12 and
14) can only be answered once the main objectives are clear.
290
(b) What information and Resources are available or obtainable?
It is no good having a superb model for a situation if in practice the

relevant data are unobtainable or unreliable (Chapter 13). Similarly, it
is no good introducing a sophisticated forecasting programme on the
computer if nobody in the firm understands its basic assumptions and
principles. The forecasting models, methods and application need to be
related to the real situation in which they are to be used.
(c) What model should be used ?

We have emphasized the concepts of structure (section 1.3) and of
models (Chapters 2 and 4 to 12) throughout this book. The problem of
identifying the model to use is basic (Chapter 12) and it must be
acknowledged that a lot depends on the experience and knowledge of
the forecaster. The more complex the models in use the more difficult
the task (e.g. Chapter 7 and Appendix A). An approach often adopted
is to use a simple model to obtain forecasts and to seek to improve it by
looking for structure in the forecast errors (Chapter 15). It is sometimes
found that a reasonable forecast can be obtained from a relatively
simple intuitively chosen model and that improvements on this are not
financially worth the effort involved. Where it is worthwhile, one can
increase the sophistication of the models or find other ways of
improving the forecasts without changing the basic model (Chapter 17).
(d) How should I regard the model?
Throughout this book we have used the concepts of global and local
models (e.g. section 2.3). These ideas emerge from the consideration of
stability as underlying the approach of statistical forecasting. If we can
assume that we have correctly identified the model, that it is stable or
that, if it does undergo a sudden change, we can identify and use the
new model rapidly, then a global use of our model is clearly the best
approach. If, however, we judge the situation to be subject to slow
evolution in time, either on the basis of past data or knowledge of the
environment in which the data will arise, then we can regard the model
as an approximation that is fitted locally for forecasting purposes (e.g.
Chapters 4 and 5).
(e) What method should be used?
Having decided on the model and the approach to its use, the
method of obtaining a forecast is usually fairly clear, as we have
illustrated in Chapters 4 to 11 and 14. A general discussion of methods
was given in Chapter 12. Often with experience we will seek to extend
and modify our initial methods (section 14.2 and chapters 17 and 18).
291
(f) How are the results of forecasting to be assessed and controlled ?

In this chapter we have emphasized the essential relatedness of the
forecast and its application. Thus to assess the forecasts we need to
look at their value in terms of the application as well as in the statistical
terms (Chapters 3, 15, 17 and 18). Often, of course, these two are
intimately related. When we turn from a general investigation of our
methods and start to develop a practical forecasting system, then we
need methods of control of the day-to-day running of the system
(Chapter 16).
References
Forrester, J. N. (1961). Industrial Dynamics. MIT Press, Cambridge, Mass.
Kuchn, A. A. and Day, R. L. (1963). The acceleration effect in forecasting
industrial shipments. Journal of Marketing, 27, 25—28.
Leicester, C. S. (1963). The cause of the shoe trade cycle. British Boot and Shoe
Institute, 11, 307—316.
Morse, P. M. (1957). Queues, Inventories and Maintainance, 1967 edn. John Wiley
and Sons, New York.
Thomeycroft, W. T., Greener, J. W. and Patrick, H. A. (1968). Investment decisions
under uncertainty and variability. Op. Res. Quarterly, 19,143—160.
Trigg, D. W. and Pitts, E. (1963). The optimal choice of the smoothing constant in
an inventory control system. Op. Res. Quarterly, 13, 287—298.
.
„
Appendix A
Some terminology and definitions

for stochastic processes
The aim of this appendix is to define some of the terms that have been
used in previous chapters, particularly Chapter 7. It also seeks to
provide some additional ideas that will be of use when some of the
more advanced references are consulted.
A discrete stochastic process is a process that generates a sequence of
observations, {xj, in time. Figure A.l represents a particular sequence
from such a process, often called a ‘realization’ of the process. The
probability density function of x* at a specific time i is given by
fixfidxi = Prob{x; < Xj < xt + dxfi
i.e. it is the probability that the sequence passes through the small
window, dxj, in an imaginary wall built at time i. Similarly, if we
imagine two such windows at times i and j, we can obtain a joint
probability density, f(xhXj), as being defined from the probability of
the process passing through both windows, dxt and dxj, as in Figure
A.l(b). If we know that at time i the sequence has the value x,, then
the probability of it passing through the window dxj is f(Xj \ xfidxj; this
is the conditional probability.
If f(Xj | X/) does not depend on the value xz-, we say that the two
random variables xt and xy are independent (Figure A.lc). If all
variables { x J at all times are independent of each other, then the whole
process is called an independent stochastic process or sometimes ‘white
noise’. The sequence{ et] used throughout this book is formally defined
as such a process.
If all the distributions of one, two or more variables depend only on
the relative positions of the ‘walls’, and thus are not effected by any
change in time origin, then the process is said to be stationary. The
global use of the models in Chapter 7 essentially assumed that the
processes being dealt with were stationary. The local use of the models
293
294
Figure A.l Illustration for probability distribution

functions of stochastic processes
295
is based on the belief that often real processes are only locally
stationary and are non-stationary over long periods of time.
We can define expectations for the above distributions as in Chapter
3, Table 3.5. Thus we can define:
E{Xi) as giving the mean value of the process at time i
E(Xj | Xj) as giving the mean value of the process at time j, given
that it took the value of xz- at time i
Var(xz) giving the variance of the process at time i
We can also define some further quantities: e.g. the autocovariance of

xt and Xj,
COv(X/,Xy) = E[{ Xt — E{Xi) K Xj ~ E(Xj)} ]
This quantity measures the relationship between xt and Xj. To see the
way in which it does this we need to modify it slightly by defining the
autocorrelation function of xt and x7 by
p(Xi,Xj) = Cov(x/,x;-)/y/(Var(x/) Var(x7))
This quantity takes the value 1 if j Xj, 1»0« if X/ and Xj are perfectly
linearly related, the value —1 if xz- = —x7 and the value 0 if xt and x7 are
independent of each other. Thus p(x,-,X/) is a measure of linear
relationship obtained by standardizing the autocovariance. Notice that
Cov(xz-,X/) = Var(x;)
If the process is stationary, Cov(x*,x7) and p(xz-,x7) will depend only on
the time difference j — i and not on the actual times. Thus putting
j = i + k we can simplify our notation for a stationary process:
Cov (xhxi+k) = C(k)
P (xi9xi+tc) = p(k)
Var<X/)= C(0)
so,
p(k) = C(k)/C( 0)
The plot of p(k) against k is called the correlogram.
By way of example, consider the first-order autoregressive process
x* = 0x,_ i + et (| 0 | < 1)
if this is re-expressed as
X, = et + 4>et_t +<A2e t„2 + . ..
296
it is straightforward to show that

E(xt) = 0
Var(X,) = (1 + 02 + </>4 + . . .)o2 = o2 1(1 - <t>2)
E(xtxt+k) = <pk(l+4>2 +<t>4 + ... )a2 = o2<t>k 1(1 — <t>2)
Hence,
Cov(xt,xt+ k) = o2<pk /(I — 4>2)
and
p(Xt,Xt+k) ~ 0
None of these depend on t.
To develop the autocorrelation idea a little further, consider a

third-order autoregressive process (see section 7.4):
+ + e
Xt — 0iX,_ i + 03X
t- 3 t
If we multiply both sides by j and take expectations, we get

C(1) = 01C(O) + 02C(1) + 03C(2)
Dividing by C(0) gives
P(1) = 01 +02P(1) + 03p(2)
If we multiply instead by xt_ 2 and repeat the process, we finish with
p(2) = 0! p(l) + 02 + 03P(1)
Finally, if we multiply by xt _ 3 instead of xt_ 2 and repeat, we obtain
+
P( 3) = 0iP(2) + 02 P( 1) 03
Thus we obtain three simultaneous equations relating the 0 and the
autocorrelations. These equations are called the Yule—Walker equations
and are the model equations that correspond to the normal equation
that we referred to in section 7.4. If we estimate p(l), p(2) and p(3)
from the data, we will obtain estimates of 0X, 02 and 03. The estimate
of 03,03 3, is of particular importance, since if 03 3 is close to zero it
suggests that the model should have order less than three. 03 3 is called
the estimated third partial autocorrelation coefficient. Thus, in general,
the quantity 4>kk may be obtained by solving the Yule—Walker
equations assuming a fcth-order autoregressive model. The sequence
0ii? 02 2 ? • • • is use<3 to decide on the order of the autoregressive
process (see Table 12.2).
A process for which E(xt) and p(xtixt+k) do not depend on t, but
which may not be totally stationary in other properties, is said to be
‘wide sense’ stationary. If a process is stationary, it will automatically
297
be wide sense stationary. The converse, though obviously not true in

general, does hold if all the distributions in the process are normal,
either univariate or multivariate. The normal distribution, sometimes
called the Gaussian distribution, plays a major role in the theory of
stochastic processes. This is because it has a number of valuable
properties in theory and occurs fairly frequently in practice. In
summary these are as follows.
(a) If the process is normal, all the distributions f(Xj,Xj), f(xt),
f{Xj | Xj)f etc., will be normal.
(b) Any linear sum of the process observations will be normal. For
example, the exponentially weighted moving average obtained by
exact exponential smoothing will be normally distributed.
(c) The three quantities E(xt) = p, Var(x,) = o2 and p(xhxi+k) = p(k)
completely define the stationary normal process.
(d) If p{k) = 0 (k = 1,2, . . .), then the process is the independent
normal process.
(e) The function of xt,xt_ i, . . . that gives the forecast with smallest
mean square error is in fact a linear combination of past
observations.
Appendix B
Proof that conditional expectation

gives minimum MSE forecasts
We here show that the forecast with the minimum mean square error is
simply the expected value of the future xt+k, given that we already
know the values OC | ^ OC 2 y * • • + ^ • This conditional expectation we
denote by
E^Xt+k I X i , . . . , Xf )
Let us start by asking what constant 5ct k we would use to forecast xt+ k
to obtain the minimum mean square error. To find this we simply
expand the mean square error and choose 5ct k to give the minimum;
thus
MSE =E(xt+k —xt'k)2
= E(xj+k) — 2xt,kE(xt+k) +x2:k
Differentiating with respect to xt k and equating to zero gives
xt,k=E(xt+k) (1)
and the mean square error for this xt k is
MSE =E{xt+k — E(xr+k)}2 = Vax(xt+k)
We thus obtain the best constant forecast by taking xt k equal to the
expected value of the future xt+k.
To obtain a realistic forecast of xt+k we must obviously use our data
xxt, which for now we regard as a sequence of random
variables. Suppose we write the forecast as
'V* /y (/V* /y* \

^ t,k t,k v**-1 ? • • • ? t)
We now want to find the function xtk (xl5 . . . , xt) which gives the
minimum mean square forecast error, so we want to minimize
298
299
MSE = E{xt+k —xttk(xl,... ,xt)}2
Now an ordinary expectation can be written in terms of conditional

expectations. If we denote an expectation with respect to a random
variable z by Ez(*), then a property of expectations is that the
expectation with respect to two variables y and z can be written as
E (• )=E{E(- | z)}
(y r-O 2 y
Thus we first hold z constant and find the expectation with respect to
y. This will give an answer which depends on z. We now find the
expectation of this quantity with respect to Applying this idea to
finding the mean square error we can rewrite the mean square error as
MSE =E[E {(xt+k — xtk(xi. ,xt)}2 I Xj,. . . , x,} ]
x Xt+k
where the inner expectation is the conditional expectation given

x
i» • • • , xt, denoted by x, and the outer expectation is with respect to
the joint distribution of x. Now conditional on OC | OC the forecast ^ • t ^ -f- y
xtk (xj, . . . , xt) is simply a constant, xtk, so the inner expectation

could be written as
E \.{.xt+k x
t,k} I x]
x
t+ k
We can choose xt k to minimize this using the result obtained above, so

the constant xtk is the expectation of xt+k, only now it is the
expectation conditional on x, i.e.
xt,k =E{xt+k | x}
This choice must obviously minimize the total mean square error
expression, as it is only the inner expectation that relates the past,
Xj, . . . , xt, to the future, xt+k. We thus arrive at the conclusion that
the minimum mean square error forecast of a future observation, xt+k,
is the expectation of that observation given the data, x1,...,xf,
available at present, i.e.
x x
t, k (*^T > • • • 5 f) ~ E {Xf + k | Xj, . . . , xt~j
Again using the results obtained for the constant forecast, it is clear
that the minimum mean square error obtained using xt^k (xx, . . . , xf)
is
MSE = E{ Var(x^+/c | xx, . . . , xt)}

X
That xtfk (Xj,. . . , x?) is unbiased follows from studying the error, which
is
@t+ k ~ Xf+k X-i k (x 1 , . . . > X^)
300
Taking conditional expectations of both sides gives
E(et+k | x) = E(xt+k | x) — E{xttk(xly . . . , xt) | x}

~ Xf k{x i, . . . > ) X t>k (*^ 1 ? • • • •> Xf )
SO
E(et+k | x) = 0
But
E(et+k) = E{E(et+k \ x)}= E (0)

X X
and so
=
E{et+ k:) 0
and
xtk (*i 5 • • • > xt) is thus an unbiased forecast of Xt

Appendix C
Two-way exponential smoothing
In the discussion of local models and discounted least squares we have

always used weights that put the greatest emphasis on present data.
From the forecasting point of view this is the logical thing to do.
However, if we wish to explore the local variation in the fitted model
over past time this is not so useful. Suppose, for example, we wish to
estimate the local values of the m odel parameters at some past time T.
In this case one logical set of weights would be
Time 0 1 2 .. . T-l T T+l • • • t
Weight aT aT~l aT~2 ... a 1 a ... a t- T

If we used these in discounted least squares to fit the model,
discounting into the future from T as well as the past, we would
minimize
t- T
X a | r | (residualr + r)2
r=-T
In general this is a more complicated exercise than we have considered

for estimating present parameter values. If we are estimating the local
mean, with large body data, the method leads to a very simple
modification of exponential smoothing. We simply generate weighted
sums by sweeping both ways through the data. Using arrows to denote
the direction of sweep, let
Si = aSi_ i + Xi = x0
St = aSi+! + Xf t
aWf_ j +1 W o =1
Wi = aWi+ i +1 Wt = 1
Then the local estimate of pT is
AT = (ST + ST -xT)l(WT +WT-1)
301
\
Appendix D
A technical note on discounted least squares
In Chapters 5 and 6 the basic concepts of discounted least squares were

introduced and some results quoted. The aim of this appendix is to
derive the normal equations for the linear form of model used in several
chapters of this book, see Gilchrist (1967). Using the notation of
Chapter 6, consider the particular model with, for the sake of
illustration, three parameters. Thus the fitted model will look like
=
yt- r b0 +b1x1 >t_r + b2X2,t-r + ^t-r
where the x values are known regressor variables, et_ r are the residuals
and t — r is the time r time units before the current time t. The
criterion for discounted least squares
S = V ae _
r=Q
r 2
t r
and we require b0, bx and b2 to be the values that minimize this with a
as the fixed discounting factor (0 < a < 1). Substituting for e gives
t- I
S= 2 ar(yt-r — b0 — £>1*1,— b2x2't_r)2

r=0
To minimize this we differentiate with respect to b0, bx and b2 and

equate each expression to zero. Thus
3s t- i
— =-2 2 ar(yt_r — bo—b1xi't_r — b2x2,t-r) =
0
Ob0 r=0
3s t- i
— =—2 2 arxitt_r(yt_r — b0—blxl t_r — b2X2,t-r) =
0
OOl r=0
35 t- I
— =—2 2 arx2tt_r(yt_r- b0 — btx, it_ r — b2x2 r) = 0

Ob2 r=0
Sorting these terms out and dropping the summation limits for clarity
302
303
gives
b0l,ar + b1'Larxltt_r + b2larx2,t-r = 2aryt_r
b0Xat'xl yt_r + biltfxl't-r + b2Ilarx1 ,t_rx2it_r
= 2arxx ,f_ ryt-r
b0T,arx2 ,f_r + SG"X2
r
jf_r + b2ô x2yt_r
= Zarx2ft_ryt_r
These are the normal equations discussed in section 6.2. Notice that the
three equations are equivalent to
'Laret_ r =0
2G 0
Y,arX2 t— r^t— r = 0
The normal equations used in section 5.2.3 are obtained by a simple
change in notation:
(a) b0 becomes the estimate, ptt, of the current mean.
(b) b1 becomes the current estimate of slope (3t by letting xt^r be
simply —r, the time into the past.
(c) x2 and b2 are set at zero as there are only two parameters and two
equations.
(d) We use x as the dependent variable instead of the y here.
We thus obtain
f.\t'Lar — jS^Za'V = Xarxt_r
pLtlLar r — r2 ='Larrxt_r
Reference
Gilchrist, W. G. (1967). Methods of estimation using discounting. J. Roy. Statist.
Soc., 29, 355-369.
X
.
Index
Acceleration effect, 279 Chatfield, C., 149

Adaptive methods, 203 Chen, G. V. C., 207
Adjustments of data, 196 Chisholme, R. K., 5,11
Adjustment of forecasts, 256 Chorofas, D. M., 168
Adulterated data, 240 Chow, W.M., 204
Albert, A. E., 208 Combining forecasts, 267
Allen, D. M., 84 Comparison of methods, 25,179, 231,
Analysis of variance, 186 240
Anderson, 0. D., 101 seasonal methods, 148
Asymptotic growth curves, 152 Computers, vii, 37,110, 256
Autocorrelation, 232, 295 Conditional distributions, 230, 293
adjustment for, 262 Conditional expectation, 100, 298
partial, 182, 296 Conditional forecast, 90
tracking signal, 254 Confidence interval (prediction interval),
Autocovariance, 295 45,63,98,104,106,241
Autoregression of errors, 232 Correlation, 224, 237
Autoregressive model, 93,105,176, local, 224
262,296 multiple, 85
Average run length, 253 Correlogram, 233, 295
Coutie, G. A., 65, 252
Ball, R. J., 90 Cowden, D. J., 61, 76
Bamber, D. J., 206 Craig, A. T., 33, 45
Barham, R. H., 157 Criteria classification, 25
Bassie, V. L., 188 cost, 172, 270
Bates, J. M., 267 practical, 275
Bayesian forecasting, 114,170, 209, probability, 170
213 statistical, 29, 34
Beale, E. M. L., 84 steady state, 32, 34
Berrisford, H. G., 207 transient, 32, 34
Bias, 30,34,44,55,62,259 Croxton, F. E., 61
Boot, J. G. C., 197 Crystal ball forecasts, 281
Box—Jenkins, 101, 104, 107, 147,148, Curley, M. L., 195
149 Cusum methods, 252
Brenner, J. E., 209 Cycles, 117
Brown, R. G., 53, 73,147, 165, 248,
249,252 Data, 12, 193
Bunn, D. W., 267 adjustment, 196
Burns, T., 90 collection, 12, 193
outliers, 201
Causal methods, 5 quality, 195
Centred moving average, 119 reduction, 12
305
306
sectioning, 158 Fisher, R. A., 46

transformations, 200 Fitting curves, 16, 58, 156
Dauten, C. A., 117 Forecast adjustments, 256
Davis, H. T., 6,11, 50,117 autocorrelation, 262
Davis, R., 240 bias, 259
Decisions, 172, 282 constraints, 264
Deseasonalized data, 135 costs, 268
D’Esopo, D. A., 209 decoupling, 274
Deterministic models, 17 probabilities, 27
Difference operator, 111 regression, 263
Discounted frequency, 165 Forecast analysis, 221
Discounted least squares, 69, 80, 99, graphical, 223
180, 302 sensitivity, 243
Discounted likelihood, 163, 172 theoretical, 33
Discounting factor (smoothing constant), Forecasting formulae, 21
51, 53 error correction form, 43, 54, 68,
adaptive, 204 131,209
choosing, 238 explicit form, 41, 51
Distributions, probability density, 293 recurrence form, 42, 53, 65,130,
errors, 226 132, 165,180,206
future observations, 229 Forecasting parameters, 22, 237
Double exponential smoothing, 72 Forecasting in a system, 278, 287
Draper, N. R., 80, 91, 157 Forecast observation diagram, 222
Durbin, J., 139 Forecasts as random variables, 281
Duvall, R. M., 140,142 Forrester, J. N., 288
Fourier smoothing, 141
Econometric forecasting, 5, 88
Eilon, S., 204 Gaussian distribution, 295
Elmaleh, J., 204 Gilchrist, W.G., 69, 302
Endogenous variables, 89 Ginsburgh, V., 197
Error analysis, 225 Godolphin, I. J., 240
correction methods, 43, 54, 68, 131, Gompertz curve, 151,154
209 Granger, C. W. J., 216, 239, 240, 269
mean, 29 Graphical analysis, 222
mean absolute, 30 Graybill, F. A., 33, 45
mean absolute deviation, 31 Gregg, J. V., 155
mean square, 30, 44, 62 Grether, D. M., 148
minimum, 298 Groff, G. K., 148, 240
root mean square, 31 Growth curves, 150
variance, 31, 34 Grunwald, H., 114
Estimation, 16,179,180
Expectation, 34 Hannan, E. J., 174
Exponential growth, 150, 154 Harrison, P. J., 114,139,141, 158, 211, 253
Exponential smoothing, 53, 108 Hill climbing methods, 157
adaptive, 203 Hogg, R. V., 33, 45
multivariate, 175 Holiday effects, 87
probability, 165 Holt, C., 65
two-way, 224, 234, 301 Holt’s method, 65, 114, 209
Exponentially weighted moving average, Holt—Winters method, 138
51
Extrapolation, 5,15 Identification, 16,179,181, 182
Information, 171
Farmer, E. D., 92 Integrated autoregressive moving average
Feibes, W., 197 model, 112
307
Intuitive methods, 4, 180 Mixed autoregressive moving average
Ivakhnenko, A. G., 92 model, 107, 177
Mixed methods, 216
Jenkins, G. M., see Box—Jenkins Models, classification, 17, 39
Jones, R. H., 176 construction, 13
Jorgenson, D. W., 139 global, 19,29,41,56, 58,124
Judgement, 186 identification, 16, 179,181
linear, 19, 78, 302
Kalman models, 209, 211 local, 19,48, 56,63,129
Kalman, R. E., 176, 212 non-linear, 19, 150
Kendall, M. G., 121, 183 testing, 183
Klein, S., 61 Montgomery, D. C., 205
Kuchn, A. A., 279 Mood, A. M., 23, 45
Morrell, J., 5, 11
Lagged variables, 79 Morrison, N., 212
Lapa, V. G., 92 Morse, P. M., 283
Leach, A. G., 206 Moving average, 49, 118, 135
Lead time, 21 centred, 119
Least squares, 60, 99, 104, 107, 180 weighted, 50,120
Leicester, C. S., 281 Moving average model, 94,101,176
Lewandowski, R., 153 Murphy, M. J., 139
Likelihood, 163 Multicolinearity, 90
Linear forecasting formulae, 213, 297 Multiple correlation, 85
Linear model, 19, 88 discounted, 85
Lisman, J. H. C., 197 Multivariate models, 174
Local bias, 234
correlation, 224 Naive forecasts, 5
mean square error, 234 Nelson, C. R., 101
models, 19, 48, 63, 290 Nerlove, M., 148
standard deviation, 234 Nettheim, N. E., 142
Logarithmic parabola, 151 154 Newbold, P.,216, 241,267
Logarithmic transformation, 118, 150, Non-linear models, 19, 150
156,200 Normal distribution, 45, 297
Logistic curve, 151, 154 plotting, 227
Long-term forecasting, 160 Normal equations, 61, 69, 80,143, 214,
Lovell, M. C., 139 296,302
McKenzie, E., 240 Overfitting, 182

Mandelbrol, B., 199
Marchant, L. J., 240 Parameters, 19
Market share, 168 forecasting, 22
Markov chains, 166 Partial autocorrelation, 182, 296
Marquardt, D. W., 157 Pearce, C., 194
Maximum likelihood, 163 Polynomial trends, 76
Mean absolute deviation, MAD, 248 Predetermined variables, 89
Mean absolute deviation of error, Prediction intervals, 45, 63, 98, 104,
31 106,241
Mean absolute error, 30, 34 Probabilistic models, 162
Mean error, 29 probability distributions, 162
Mean square error, 31, 34, 35, 45, 62, state models, 166
97,233,237,239 Probability distributions, 14, 293
Mean square successive difference, 254 Prothero, D. L., 149
Minimum mean square error, 96, 100,
297,298 Quality control of forecasts, 246
308
Quality of data, 195 Stationary process, 293
Queues, 283 Statistical equilibrium, 246
Steady state, 29, 32
Rand Corporation, 37 Steps in forecasting, 12
Random variables, 14, 281 Stochastic models, 19, 93, 147, 176,
Rao, A. G., 206 293
Recurrence forecasting formulae, 42, Stockholding effects, 280
53, 65,108,130,132, 138,165, Structure, mathematical, 7, 13
180,206 statistical, 1, 6, 14, 246
Recurrence form models, 210 Stuart, A., 121
Reed, R., 148, 205
Regression models, 77 ^-distribution, 47
dependent variables, 79 Tests, randomness, 184
fitting, 79 seasonality, 186
on time, 90 trend, 185
regressor variables, 79, 86, 90, 92 Theil, H., 171, 225
selecting, 83 Theil’s coefficients, 235
Reid, D. J., 148 Thorneycroft, W. T., 287
Residuals, 60, 69, 80, 123 Tier chart, 122
Roberts, S. D., 205 Time-series, v (see also Stochastic
Robinson, C., 5,11, 92 methods)
Robinson, E., 174, 215 Tracking signal, bias, 248
autocorrelation, 254
Safety stock, 275 Transformations, 156
Sample variance, 31 Transients, 28, 32
Scatter diagrams, 230 Transition probabilities, 166
Scheduling, 281, 283 Trends, linear, 150
Seasonal adjustment, 135 non-linear growth, 58
Seasonal forecasting, 115 polynomial, 75
Fourier methods, 139 Trigg, D. W., 206, 248, 275
index methods, 124, 210 Turning points, 184
stochastic methods, 147 Two-stage forecasting, 256
Seasonal index, 135
Second stage forecast, 257 Van Dobbin de Bruyn, C. S., 205
Sensitivity analysis, 243 Variance, 34
Self-defeating forecasts, 287 Von-Neuman ratio, 254
Self-validating forecasts, 287
Short-term forecasting, 160
Wagle, B.,241
Silk, L. S., 195
Ward, D. H., 209
Simple modified exponential growth,
Weighted moving average, 50, 120
151, 154, 158, 161, 210
Weiner, N., 215
Simulation, 36, 94
Wheelwright, S. C., 216
Singleton, P., 85
Whittle, P., 215
Slope characteristics, 153
Whybark, D. C., 205
Slutsky effect, 120
Williams, W. H., 242, 243
Smith, H., 80, 91
Winters, P. R., 53, 138, 207
Smoothing, 55, 72
Sources of data, 12, 193 Woy, J. B., 87,195
Spencer, M. H., 91
Stability, 6, 246 Yaglom, A. N., 215
Stagewise least squares, 91 Yates, F., 46
Starting values, 65 Yule—Walker equations, 296
State models, 166
Stationarity, 293 Zeilner, A., 266
220709
QA Gilchrist
280
. G54 Statistical forecasting
DATE DUE
mi i tj
Lebanon Valley College Library

Annville, Pennsylvania 17003

Statistical Forecasting

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Statistical Forecasting

Uploaded by

Copyright:

Available Formats

Statistical Forecasting

JOHN WILEY & SONS

Reprinted September 1977

All rights reserved.

No part of this book may be reproduced by any means, nor

Library of Congress Cataloging in Publication Data:

Gilchrist, Warren Garland, 1932—

ISBN 0 471 99402 2 (Cloth)

Typeset in IBM Century by

Figure P.l The structure of the book and of statistical

trending time series, another seasonally varying time series. We then

This interrelation between forecast and application is explored in

I would like to express my thanks to the many people who have

Chapter 1. An introduction to forecasting 3

Chapter 3. Forecasting criteria 23

PART II FORECASTING FOR SOME BASIC MODELS

Chapter 4. The constant mean model 41

Chapter 5. Linear trend models 58

Chapter 6. Regression models 77

Chapter 7. Stochastic models 93

Chapter 8. Seasonal models 115

Chapter 9. Growth curves 150

Chapter 10. Probabilistic models 162

Chapter 11. Multivariate models 174

Chapter 12. Forecasting methods and models 179

PART III THE FORECASTING PROCESS

Chapter 13. Data 193

Chapter 14. Adaptive methods and other extensions 203

Chapter 15. The analysis and comparison of methods 220

Chapter 16. Forecast control 246

Chapter 17. Two-stage forecasting 256

Chapter 18. Problems in practice 274

Appendix A Some terminology and definitions for stochastic

Appendix B Proof that conditional expectation gives minimum

Appendix C Two-way exponential smoothing 301

Appendix D A technical note on discounted least squares . . . .302

1.1 Everybody forecasts

What will next month’s sales be?

To obtain the best anwers to these questions, we would need to be able

1.2 Methods of forecasting

1.2.1 Intuitive methods

Alternatively, a sample survey of the views of potential customers

1.2.2 Causal methods

1.2.3 Extrapolative methods

1.3 The basis of forecasting

the structure of society as they see it and take action to introduce

As a third example we will consider some of the forms taken by a set of

week, week 8, but no longer with such confidence. Suppose that we

Figure 1.1 Bar chart of the data of

The structure of this situation is thus a structure in the probabilities

Figure 1.2 Plot of the data of Table 1.1(e)

week 13 would be our forecast of 6,000 units as made for week 13 in

Table 1.2 Two years’ data

Figure 1.3 Plot of the data of Table 1.2

Table 1.3 Breakdown of data in Table 1.2

(b) Frequency table for random variation

Values Less than —7, --7 to —5, —4 to —2, —1 to 1, 2 to 4, 5 to 7, 8 to 10

2.1 Scientific forecasting

It has been seen in Chapter 1 that forecasting is an inherent part of

2.1.1 The first step — data collection

Before we can apply any scientific forecasting method to a situation,

2.1.2 The second step — data reduction

are to be included in this basic minimum information three criteria are

Relevance: Is the information the most directly relevant information